Spaces:
Sleeping
Sleeping
| json_file,model_name,model_private,model_num_parameters,assin2_rte,assin2_sts,faquad_nli,hatebr_offensive | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-1.5-34B/results_2024-05-17T10-36-18.336343.json,01-ai/Yi-1.5-34B,False,34388917248,0.8976911637262349,0.8148786802023537,0.585644163957417,0.8363023241432246 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-1.5-34B-32K/results_2024-05-20T21-45-47.139761.json,01-ai/Yi-1.5-34B-32K,False,34388917248,0.9121699049758872,0.809949940837174,0.7177866756717641,0.8271604938271605 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-1.5-34B-Chat/results_2024-05-16T15-15-24.863291.json,01-ai/Yi-1.5-34B-Chat,False,34388917248,0.9170744853491483,0.7661887019644651,0.7743940809133725,0.8210886883714428 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-1.5-34B-Chat-16K/results_2024-05-20T18-32-22.664525.json,01-ai/Yi-1.5-34B-Chat-16K,False,34388917248,0.9116655919331504,0.777225956509937,0.7909900023305279,0.8889320535439632 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-1.5-6B/results_2024-05-17T13-51-05.776238.json,01-ai/Yi-1.5-6B,False,6061035520,0.85320443811568,0.611946662194731,0.566892243623113,0.8390372896945542 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-1.5-6B-Chat/results_2024-05-17T14-53-37.626126.json,01-ai/Yi-1.5-6B-Chat,False,6061035520,0.8478217777818736,0.6797897994537765,0.6548247706694055,0.7881170986195587 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-1.5-9B/results_2024-05-17T09-09-38.931019.json,01-ai/Yi-1.5-9B,False,8829407232,0.8815204475360152,0.7102876692830821,0.6362495548508539,0.7837384886240519 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-1.5-9B-32K/results_2024-05-20T20-18-06.355522.json,01-ai/Yi-1.5-9B-32K,False,8829407232,0.8657419886018202,0.7267527969011244,0.5410839160839161,0.7806530019415174 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-1.5-9B-Chat/results_2024-05-17T02-43-33.664147.json,01-ai/Yi-1.5-9B-Chat,False,8829407232,0.8807093197512774,0.7520700202607307,0.6913654763916721,0.8297877646706737 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-1.5-9B-Chat-16K/results_2024-05-20T09-41-40.204030.json,01-ai/Yi-1.5-9B-Chat-16K,False,8829407232,0.889630445004916,0.7254379491320617,0.6373099047367834,0.8668983847883869 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-34B/results_2024-02-07T03-32-43.048071.json,01-ai/Yi-34B,False,34388917248,0.7095337812960236,0.6212032386293976,0.7969022005981341,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-34B/results_2024-04-13T15-53-49.411062.json,01-ai/Yi-34B,False,34388917248,0.917882167398896,0.76681855136608,0.7798334442926054,0.8107834570679608 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-34B-200K/results_2024-02-09T08-03-26.811325.json,01-ai/Yi-34B-200K,False,34388917248,0.7858403678133732,0.5583683246827316,0.7800338409475465,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-34B-200K/results_2024-04-17T23-49-34.862700.json,01-ai/Yi-34B-200K,False,34388917248,0.9097218456052794,0.7390390977418284,0.49676238738738737,0.8117947554592124 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-34B-Chat/results_2024-02-28T08-14-36.046639.json,01-ai/Yi-34B-Chat,False,34388917248,0.924014535978148,0.7419038025688336,0.7157210401891253,0.7198401711140126 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-6B/results_2024-02-06T22-20-36.837117.json,01-ai/Yi-6B,False,6061035520,0.48362346630826814,0.44473899241461606,0.6078037007240547,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-6B/results_2024-02-17T03-42-08.504508.json,01-ai/Yi-6B,False,6061035520,0.7903932929806128,0.5666878345297481,0.5985418799210473,0.7425595238095237 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-6B-200K/results_2024-02-09T07-10-30.923872.json,01-ai/Yi-6B-200K,False,6061035520,0.7574463542222696,0.3059276997249324,0.4471267110923455,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-6B-200K/results_2024-04-16T17-07-31.622853.json,01-ai/Yi-6B-200K,False,6061035520,0.40523403335417163,0.4964641013268987,0.4396551724137931,0.4892942520605069 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-6B-Chat/results_2024-02-28T14-35-07.615539.json,01-ai/Yi-6B-Chat,False,6061035520,0.7948490568935549,0.5684271643349206,0.637960088691796,0.775686136523575 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-9B/results_2024-04-14T11-08-02.891090.json,01-ai/Yi-9B,False,8829407232,0.8784695970900473,0.752860308487488,0.7478708154144531,0.8574531631821884 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-9B-200K/results_2024-04-14T12-49-52.148781.json,01-ai/Yi-9B-200k,False,8829407232,0.8708321784112503,0.7508245525986388,0.7162112665738773,0.8238294119604646 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-9B-200K/results_2024-05-20T06-21-38.524751.json,01-ai/Yi-9B-200K,False,8829407232,0.8735592777805543,0.7486645258696737,0.7445188998494914,0.817858599988343 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-Coder-9B/results_2024-10-02T03-06-38.881073.json,01-ai/Yi-Coder-9B,False,8829407232,0.8479141293428532,0.7388500902557162,0.5212030552344689,0.8155218925088192 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/01-ai/Yi-Coder-9B-Chat/results_2024-10-02T01-59-56.934513.json,01-ai/Yi-Coder-9B-Chat,False,8829407232,0.7994152126314071,0.7640954232319354,0.6541275915011303,0.8464002808680579 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/152334H/miqu-1-70b-sf/results_2024-05-23T11-20-45.843993.json,152334H/miqu-1-70b-sf,False,68976648192,0.9309631191550011,0.6972270341129023,0.7641750093536355,0.8382584367896485 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/22h/cabrita_7b_pt_850000/results_2024-02-16T13-19-48.657595.json,22h/cabrita_7b_pt_850000,False,6902255616,0.6996743545023234,0.007646758869425693,0.17721518987341772,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/22h/cabrita_7b_pt_850000/results_2024-03-08T02-07-35.059732.json,22h/cabrita_7b_pt_850000,False,6902255616,0.3333333333333333,0.1265472264440735,0.17721518987341772,0.5597546967409981 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/22h/open-cabrita3b/results_2024-02-16T12-41-51.764611.json,22h/open-cabrita3b,False,3554473600,0.38392470950636587,0.037061970342522815,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/22h/open-cabrita3b/results_2024-02-28T16-38-27.766897.json,22h/open-cabrita3b,False,3554473600,0.4301327637723658,0.08919111846797594,0.4396551724137931,0.5046251022011318 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AALF/gemma-2-27b-it-SimPO-37K-100steps/results_2024-10-02T01-48-29.895681.json,AALF/gemma-2-27b-it-SimPO-37K-100steps,False,27227128320,0.8957539543945432,0.7116783323256065,0.7043137254901961,0.8798753128354102 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/abacusai/bigstral-12b-32k/results_2025-02-05T04-46-15.076283.json,abacusai/bigstral-12b-32k,False,12476420096,0.9072277560791967,0.7306032859296363,0.5912319890525638,0.8250337512054002 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/abacusai/bigyi-15b/results_2025-02-05T14-06-54.358143.json,abacusai/bigyi-15b,False,15058243584,0.795611007144068,0.6803659917540784,0.5721293440819493,0.8753272914495813 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/abacusai/Liberated-Qwen1.5-14B/results_2024-06-22T02-39-55.266663.json,abacusai/Liberated-Qwen1.5-14B,False,14167290880,0.925217532107113,0.7986766597023826,0.446593203434771,0.8469486355683289 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/abacusai/Llama-3-Smaug-8B/results_2024-05-17T21-40-04.632269.json,abacusai/Llama-3-Smaug-8B,False,8030261248,0.8853889312602414,0.7031337316356315,0.5615099394315888,0.8806396855177343 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/abacusai/Slerp-CM-mist-dpo/results_2025-02-04T09-48-47.460109.json,abacusai/Slerp-CM-mist-dpo,False,7241732096,0.924004948303844,0.7855680177965118,0.7805128205128204,0.8437745393828681 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/abacusai/Smaug-34B-v0.1/results_2024-04-18T19-26-05.567588.json,abacusai/Smaug-34B-v0.1,False,34388917248,0.918703167734794,0.8085686681843944,0.8052119294207902,0.6963199072732542 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/abacusai/Smaug-72B-v0.1/results_2024-03-06T04-23-36.061378.json,abacusai/Smaug-72B-v0.1,False,72288575488,0.9295578224919228,0.8178709164364611,0.7746799251702434,0.8565051020408163 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/abacusai/Smaug-Llama-3-70B-Instruct/results_2024-05-23T22-23-24.773490.json,abacusai/Smaug-Llama-3-70B-Instruct,False,70553706496,0.9362097477374545,0.7851681882574679,0.8001270220794261,0.9177933488181389 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/abhishek/autotrain-llama3-orpo-v2/results_2024-04-27T05-52-38.138924.json,abhishek/autotrain-llama3-orpo-v2,False,8030326784,0.3206194436478348,0.11473255710076143,0.0,0.329951690821256 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/adalbertojunior/Llama-3-8B-Dolphin-Portuguese/results_2024-04-27T22-18-08.276196.json,adalbertojunior/Llama-3-8B-Dolphin-Portuguese,False,8030261248,0.9284385454315845,0.7592009082968683,0.7966764418377321,0.8803890782643218 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/adalbertojunior/Llama-3-8B-Dolphin-Portuguese-v0.2/results_2024-05-01T00-46-52.845555.json,adalbertojunior/Llama-3-8B-Dolphin-Portuguese-v0.2,False,8030261248,0.9296707760641184,0.7599174826851254,0.8017064073924638,0.882563551736484 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/adalbertojunior/Llama-3-8B-Dolphin-Portuguese-v0.3/results_2024-05-21T00-42-47.332922.json,adalbertojunior/Llama-3-8B-Dolphin-Portuguese-v0.3,False,8114147328,0.9305138130845616,0.7648432769959327,0.7677867540881239,0.8324636712326505 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/adalbertojunior/Llama-3-8B-Instruct-Portuguese-v0.1/results_2024-04-20T16-07-27.469361.json,adalbertojunior/Llama-3-8B-Instruct-Portuguese-v0.1,False,8030261248,0.8979843685726039,0.7348666722148917,0.732803507810359,0.8820752294294207 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/adalbertojunior/Llama-3-8B-Instruct-Portuguese-v0.2/results_2024-04-22T19-02-02.927179.json,adalbertojunior/Llama-3-8B-Instruct-Portuguese-v0.2,False,8030277632,0.9203334459840036,0.7844428141792892,0.7603096357001466,0.8712435987876552 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/adalbertojunior/Llama-3-8B-Instruct-Portuguese-v0.2-fft/results_2024-04-22T23-44-45.695403.json,adalbertojunior/Llama-3-8B-Instruct-Portuguese-v0.2-fft,False,8030277632,0.9154394690903739,0.7789241548512724,0.6851214766037457,0.822700856945858 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/adalbertojunior/Llama-3-8B-Instruct-Portuguese-v0.3/results_2024-04-23T19-01-24.318718.json,adalbertojunior/Llama-3-8B-Instruct-Portuguese-v0.3,False,8030261248,0.895774051058968,0.7564562960862433,0.4471267110923455,0.821464409954509 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/adalbertojunior/Llama-3-8B-Instruct-Portuguese-v0.4/results_2024-04-24T04-20-43.795920.json,adalbertojunior/Llama-3-8B-Instruct-Portuguese-v0.4,False,8030261248,0.9064486128380835,0.7389090235338851,0.4396551724137931,0.6373674355957164 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/adalbertojunior/Mistral-7b-Portuguese-V0.1/results_2024-04-26T13-21-33.541008.json,adalbertojunior/Mistral-7b-Portuguese-V0.1,False,7241756672,0.3333333333333333,0.0013300802451966486,0.0,0.36904107740212666 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/adalbertojunior/NeuralDaredevil-Dolphin-Portuguese/results_2024-06-11T15-31-29.072829.json,adalbertojunior/NeuralDaredevil-Dolphin-Portuguese,False,8030261248,0.9305019305019305,0.7464639123068089,0.7775191675794085,0.8617010767559024 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/adalbertojunior/Qwen1.5-32B-Dolphin-Portuguese-v0.1/results_2024-04-30T22-20-08.456711.json,adalbertojunior/Qwen1.5-32B-Dolphin-Portuguese-v0.1,False,32512545792,0.9365953910677771,0.7770083928372574,0.821448192506318,0.8670860473371804 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AdaptLLM/finance-LLM/results_2024-04-24T13-38-50.219195.json,AdaptLLM/finance-LLM,False,6738423808,0.7173994459883221,0.3141019003448064,0.6856866537717602,0.6665618718263835 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AdaptLLM/finance-LLM-13B/results_2024-04-24T18-00-42.073230.json,AdaptLLM/finance-LLM-13B,False,13015874560,0.8704914563684142,0.6914158506759536,0.6137142857142857,0.8210157972117231 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AdaptLLM/law-LLM/results_2024-04-24T01-23-04.736612.json,AdaptLLM/law-LLM,False,6738423808,0.7690457097032879,0.2736321836385559,0.6837598520969155,0.6310564282443625 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AdaptLLM/law-LLM-13B/results_2024-04-24T17-20-32.289644.json,AdaptLLM/law-LLM-13B,False,13015874560,0.7762008093366958,0.6862803522831282,0.5589431210148192,0.7648719048333295 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AdaptLLM/medicine-LLM/results_2024-04-24T13-09-44.649718.json,AdaptLLM/medicine-LLM,False,6738423808,0.7412241742464284,0.30610797857979344,0.6385993049986635,0.4569817890542286 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AdaptLLM/medicine-LLM-13B/results_2024-04-24T17-45-23.659613.json,AdaptLLM/medicine-LLM-13B,False,13015874560,0.802953910231819,0.6774179667769704,0.7227569273678784,0.8155967923139503 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AetherResearch/Cerebrum-1.0-7b/results_2024-04-01T22-58-48.098123.json,AetherResearch/Cerebrum-1.0-7b,False,7241732096,0.8562832789419443,0.7083110279713039,0.7709976024119299,0.7925948726646638 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ai-forever/mGPT/results_2024-06-13T13-19-28.656765.json,ai-forever/mGPT,False,1417596928,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ai-forever/mGPT-13B/results_2024-06-17T01-34-12.100171.json,ai-forever/mGPT-13B,False,13108070400,0.0,0.005823464030093911,0.0,0.16181562378391323 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AI-Sweden-Models/gpt-sw3-20b/results_2024-02-08T16-32-05.080295.json,AI-Sweden-Models/gpt-sw3-20b,False,20340658176,0.5097556100727492,0.04176979032426292,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AI-Sweden-Models/gpt-sw3-20b/results_2024-06-17T02-29-40.078292.json,AI-Sweden-Models/gpt-sw3-20b,False,20340658176,0.5120595126338128,0.07348005953132232,0.4396551724137931,0.533175653895298 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AI-Sweden-Models/gpt-sw3-40b/results_2024-02-08T16-51-46.875316.json,AI-Sweden-Models/gpt-sw3-40b,False,39200899072,0.5670142345370424,0.03118303336610547,0.43118086176220316,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AI-Sweden-Models/gpt-sw3-40b/results_2024-02-21T07-59-22.606213.json,AI-Sweden-Models/gpt-sw3-40b,False,39200899072,0.4096747911636189,0.17308746611294112,0.5125406216148655,0.3920230910522173 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AI-Sweden-Models/gpt-sw3-6.7b/results_2024-02-08T14-27-00.255851.json,AI-Sweden-Models/gpt-sw3-6.7b,False,6714695680,0.3333333333333333,0.018735799862626468,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AI-Sweden-Models/gpt-sw3-6.7b/results_2024-04-15T22-34-55.424388.json,AI-Sweden-Models/gpt-sw3-6.7b,False,6714695680,0.6192900448928588,0.08103924791097977,0.4396551724137931,0.40737531518832293 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AI-Sweden-Models/gpt-sw3-6.7b-v2/results_2024-02-08T15-37-30.575084.json,AI-Sweden-Models/gpt-sw3-6.7b-v2,False,6714695680,0.3333333333333333,0.015469193291869708,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AI-Sweden-Models/gpt-sw3-6.7b-v2/results_2024-04-16T00-18-50.805343.json,AI-Sweden-Models/gpt-sw3-6.7b-v2,False,6714695680,0.5833175952742944,0.14706689693418745,0.4396551724137931,0.5569200631123247 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ai21labs/Jamba-v0.1/results_2024-05-20T11-27-25.699898.json,ai21labs/Jamba-v0.1,False,51570323328,0.8932135005194366,0.7608377349455587,0.5681153506087193,0.612922813715208 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AIDC-AI/Marco-LLM-ES/results_2025-04-06T01-12-18.981578.json,AIDC-AI/Marco-LLM-ES,False,7615616512,0.9205062830085595,0.7418390500261379,0.7559457143682347,0.8373430424906447 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allenai/OLMo-1B/results_2024-02-22T01-28-13.218663.json,allenai/OLMo-1B,False,1176764416,0.3411570175245718,0.09275261899111728,0.4396551724137931,0.41331316520851613 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allenai/OLMo-2-0325-32B/results_2025-07-07T18-40-13.005395.json,allenai/OLMo-2-0325-32B,False,32234279936,0.9168146288836421,0.0,0.7777113610042217,0.89186473987434 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allenai/OLMo-2-0325-32B-Instruct/results_2025-07-07T16-23-21.519592.json,allenai/OLMo-2-0325-32B-Instruct,False,32234279936,0.9313560085858401,0.7251473879787069,0.8051167407485031,0.8997044351108163 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allenai/OLMo-2-0425-1B/results_2025-05-04T15-32-02.154020.json,allenai/OLMo-2-0425-1B,False,1484916736,0.3582588385476761,0.025509828684116245,0.4396551724137931,0.3330157217722725 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allenai/OLMo-2-0425-1B-Instruct/results_2025-05-04T15-03-51.718500.json,allenai/OLMo-2-0425-1B-Instruct,False,1484916736,0.7702720764234168,0.37210622434136553,0.4148306088604596,0.7379162191192267 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allenai/OLMo-2-1124-7B/results_2025-05-05T06-45-19.797307.json,allenai/OLMo-2-1124-7B,False,7298617344,0.8671203991732603,0.7240236139466845,0.5254972109378048,0.6764444444444444 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allenai/OLMo-2-1124-7B-Instruct/results_2025-05-05T05-13-41.340824.json,allenai/OLMo-2-1124-7B-Instruct,False,7298617344,0.8707700326429424,0.7249106871531545,0.7055751222533441,0.772141196364405 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allenai/OLMo-7B/results_2024-02-22T07-00-35.468626.json,allenai/OLMo-7B,False,6888095744,0.33965642015109887,0.06149125456230655,0.4396551724137931,0.7262187865364063 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allenai/OLMo-7B-Twin-2T/results_2024-02-24T18-18-57.074671.json,allenai/OLMo-7B-Twin-2T,False,6888095744,0.7768479511684292,0.3105859111360108,0.4396551724137931,0.747126388216977 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allenai/tulu-2-13b/results_2024-05-22T22-52-46.662926.json,allenai/tulu-2-13b,False,13015864320,0.8574206250038096,0.7243471442167476,0.6181838094005436,0.8106498040953742 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allenai/tulu-2-7b/results_2024-05-22T22-31-06.551741.json,allenai/tulu-2-7b,False,6738415616,0.8793862084803441,0.6624516338838828,0.643738010413812,0.860702274839004 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allenai/tulu-2-dpo-13b/results_2024-03-08T19-15-36.234980.json,allenai/tulu-2-dpo-13b,False,13015864320,0.8989360181838149,0.7570093000021878,0.7264633552951698,0.7441058326667183 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allenai/tulu-2-dpo-70b/results_2024-05-22T02-18-14.364873.json,allenai/tulu-2-dpo-70b,False,68976648192,0.9317329158205281,0.7834390660894048,0.8339739223508806,0.8236992331909445 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allenai/tulu-2-dpo-7b/results_2024-04-02T00-29-33.604027.json,allenai/tulu-2-dpo-7b,False,6738415616,0.882693297611393,0.6797107988195017,0.6767812942008486,0.8485191831468404 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allknowingroger/MultiverseEx26-7B-slerp/results_2024-07-04T03-30-47.209441.json,allknowingroger/MultiverseEx26-7B-slerp,False,7241732096,0.9223768647276014,0.7803062713375329,0.7842076261469852,0.8163838486760461 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allknowingroger/Ph3della5-14B/results_2024-10-13T02-33-09.735947.json,allknowingroger/Ph3della5-14B,False,13960238080,0.9313195548489666,0.7150062521951339,0.6866859299725274,0.8461398095617265 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allknowingroger/Qwen2.5-slerp-14B/results_2024-11-02T04-11-15.780739.json,allknowingroger/Qwen2.5-slerp-14B,False,14770033664,0.9456698530460836,0.8409511123033333,0.7968815254182839,0.8880227399974159 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/allknowingroger/QwenSlerp6-14B/results_2024-12-22T03-19-05.521173.json,allknowingroger/QwenSlerp6-14B,False,14765947904,0.9477107081041861,0.8262422786178116,0.7871549565017302,0.9047947744500262 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/alpindale/WizardLM-2-8x22B/results_2024-05-24T04-13-39.697871.json,alpindale/WizardLM-2-8x22B,False,140620634112,0.9411764313247785,0.7931661141397623,0.7697292261172526,0.8206103536591043 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/alpindale/WizardLM-2-8x22B/results_2024-06-16T16-33-36.333906.json,alpindale/WizardLM-2-8x22B,False,140620634112,0.9391095909505747,0.7693957002357589,0.7720584315936833,0.8071909167367536 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/altomek/YiSM-34B-0rn/results_2024-12-07T16-05-56.837581.json,altomek/YiSM-34B-0rn,False,34388917248,0.9191020417428017,0.8058213777167098,0.8168287028141772,0.8577598989156642 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/alvarobartt/mistral-7b-orpo-alignment-handbook/results_2024-05-10T00-56-03.142744.json,alvarobartt/mistral-7b-orpo-alignment-handbook,False,7241748480,0.7654491359499884,0.5848324837621057,0.45449901481427424,0.5005847953216375 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/alvarobartt/mistral-7b-orpo-capybara-reproduction/results_2024-05-06T14-49-34.852021.json,alvarobartt/mistral-7b-orpo-capybara-reproduction,False,7241748480,0.8762219995377583,0.6901930492516065,0.7116824944372537,0.7796813115962051 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/alvarobartt/Mistral-7B-v0.1-ORPO/results_2024-05-20T23-25-31.585016.json,alvarobartt/Mistral-7B-v0.1-ORPO,False,7241748480,0.8628531861903126,0.7185765442633426,0.5064708415493756,0.8083461375004086 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/alvarobartt/mistral-orpo-mix/results_2024-05-09T01-52-18.616759.json,alvarobartt/mistral-orpo-mix,False,7241748480,0.8306576402321084,0.5664152925187476,0.6705985146352119,0.7530267389366248 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/alvarobartt/mistral-orpo-mix-b0.1-l2048-pl1792-lr5e-6-inverse-sqrt/results_2024-05-09T13-11-42.910831.json,alvarobartt/mistral-orpo-mix-b0.1-l2048-pl1792-lr5e-6-inverse-sqrt,False,7241748480,0.8240257570695259,0.5640429954912689,0.6763807753584705,0.6905952412784147 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/amazingvince/Not-WizardLM-2-7B/results_2024-05-22T01-29-17.284772.json,amazingvince/Not-WizardLM-2-7B,False,7241732096,0.8811187025669601,0.7451048378508254,0.6155496155496156,0.8299639583037598 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/amazon/MegaBeam-Mistral-7B-300k/results_2024-10-11T07-07-35.929134.json,amazon/MegaBeam-Mistral-7B-300k,False,7241732096,0.8986629811014633,0.7331743778715074,0.6124343313887904,0.7779370990971557 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/anthracite-org/magnum-v2.5-12b-kto/results_2024-10-19T03-02-06.506908.json,anthracite-org/magnum-v2.5-12b-kto,False,12247782400,0.9113400977866568,0.7590331658293094,0.5569285468892294,0.7744749127343797 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/anthracite-org/magnum-v2.5-12b-kto/results_2025-05-07T19-17-25.777956.json,anthracite-org/magnum-v2.5-12b-kto,False,0,0.8674910385642457,0.6823897405705615,0.6125554125554126,0.7709540048143357 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/anthracite-org/magnum-v3-27b-kto/results_2024-09-19T02-53-47.503956.json,anthracite-org/magnum-v3-27b-kto,False,27227128320,0.9235467994775257,0.802359277171842,0.7360672268907562,0.7596886827656057 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/anthracite-org/magnum-v3-9b-chatml/results_2024-10-19T01-39-01.153008.json,anthracite-org/magnum-v3-9b-chatml,False,9241713152,0.8787839045003027,0.7869756583256097,0.5128891393153965,0.7597560995167858 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/anthracite-org/magnum-v3-9b-customgemma2/results_2024-09-22T08-12-13.485613.json,anthracite-org/magnum-v3-9b-customgemma2,False,9241705984,0.8805032792322425,0.788601954301218,0.7058092685610834,0.7248474111756376 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/anthracite-org/magnum-v4-9b/results_2024-10-27T02-44-28.675383.json,anthracite-org/magnum-v4-9b,False,9241705984,0.35799185126211774,0.12685415173761294,0.4396551724137931,0.4400501560461546 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/arcee-ai/Arcee-Scribe/results_2024-08-12T00-20-50.024216.json,arcee-ai/Arcee-Scribe,False,7737708544,0.8998640109993313,0.799391039043716,0.7338973799126638,0.8648841174497106 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/arcee-ai/Arcee-Spark/results_2024-08-12T00-09-42.710641.json,arcee-ai/Arcee-Spark,False,7615616512,0.938716286447287,0.7631212166257828,0.7306463397345722,0.7633469292424382 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/arcee-ai/Llama-3.1-SuperNova-Lite/results_2024-09-22T01-40-41.855795.json,arcee-ai/Llama-3.1-SuperNova-Lite,False,8030261248,0.9268688219340129,0.7502529164855571,0.7163978826570514,0.8676377321386318 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/arcee-ai/SuperNova-Medius/results_2024-12-08T22-10-12.640809.json,arcee-ai/SuperNova-Medius,False,14770033664,0.946483704287288,0.8289512489583748,0.7891175380051909,0.9061872908843613 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/arcee-ai/Virtuoso-Lite/results_2025-02-05T20-52-15.309801.json,arcee-ai/Virtuoso-Lite,False,10305653760,0.935844334173138,0.8271075086247555,0.8400858124132773,0.7299751429766448 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/arcee-ai/Virtuoso-Small/results_2024-12-21T19-58-51.584611.json,arcee-ai/Virtuoso-Small,False,14770033664,0.9410932716602275,0.8374099871698568,0.7654381425274495,0.9018459411147293 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/argilla/CapybaraHermes-2.5-Mistral-7B/results_2024-03-08T11-45-00.125851.json,argilla/CapybaraHermes-2.5-Mistral-7B,False,7241748480,0.9010484174855619,0.7017862358256567,0.7489689797382105,0.7959179423530661 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/argilla/notus-7b-v1/results_2024-03-08T21-39-01.296165.json,argilla/notus-7b-v1,False,7241732096,0.8831337795605724,0.7200608035095787,0.6869493943899179,0.7412705678079557 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/argilla/notux-8x7b-v1/results_2024-03-07T15-39-25.432269.json,argilla/notux-8x7b-v1,False,46702792704,0.9265969804496399,0.8239880630155871,0.7984830206980806,0.7790834886956947 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/arthrod/cicerollamatry8/results_2024-05-29T22-31-18.219408.json,arthrod/cicerollamatry8,False,8030269440,0.9035862373409591,0.7644344968704779,0.7599145087660789,0.8526213369963369 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/artificialguybr/llama3-8b-alpacadata-ptbr/results_2024-05-28T02-42-18.113984.json,artificialguybr/llama3-8b-alpacadata-ptbr,False,8030261248,0.6025642021422405,0.017371246441167967,0.4396551724137931,0.33459895328341466 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/artificialguybr/OpenHermesV2-PTBR/results_2024-05-28T05-33-55.304080.json,artificialguybr/OpenHermesV2-PTBR,False,7241748480,0.8888882214065588,0.6761074691926932,0.5306445300657594,0.7476926315722321 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/automerger/YamshadowExperiment28-7B/results_2024-06-18T09-27-09.546862.json,automerger/YamshadowExperiment28-7B,False,7241732096,0.9223814244089994,0.778134909825148,0.7859986232079255,0.8186191231446596 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/aws-prototyping/MegaBeam-Mistral-7B-512k/results_2024-10-11T08-48-24.038535.json,aws-prototyping/MegaBeam-Mistral-7B-512k,False,7241732096,0.89367730071844,0.7065637552288675,0.4794674194843864,0.863375623533932 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/AXCXEPT/EZO-Qwen2.5-32B-Instruct/results_2024-12-03T10-35-22.020429.json,AXCXEPT/EZO-Qwen2.5-32B-Instruct,False,32763876352,0.9489341505871822,0.7368658525162626,0.8287377450980392,0.9185049019607843 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/axolotl-ai-co/romulus-mistral-nemo-12b-simpo/results_2024-09-01T06-34-38.572809.json,axolotl-ai-co/romulus-mistral-nemo-12b-simpo,False,12247782400,0.9051636192879993,0.7869752179868569,0.6804733727810651,0.844157646190224 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/axolotl-ai-co/romulus-mistral-nemo-12b-simpo/results_2024-09-24T04-25-29.602548.json,axolotl-ai-co/romulus-mistral-nemo-12b-simpo,False,12247782400,0.9051636192879993,0.7869752179868569,0.6804733727810651,0.844157646190224 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Azure99/blossom-v5.1-34b/results_2024-06-13T23-01-25.469438.json,Azure99/blossom-v5.1-34b,False,34388917248,0.9087560002226345,0.8294159674038925,0.8188429839812075,0.8519995605589193 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Aquila-7B/results_2024-02-13T21-53-04.495865.json,BAAI/Aquila-7B,False,7295537152,0.48191507425133256,0.2947702523754177,0.47700641417918904,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Aquila-7B/results_2024-04-03T05-32-42.254781.json,BAAI/Aquila-7B,False,7295537152,0.7202499022958302,0.04640761012170769,0.47034320848362593,0.6981236353283272 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Aquila2-34B/results_2024-02-12T19-17-32.271067.json,BAAI/Aquila2-34B,False,33693554688,0.7334164728077959,0.4958376987221018,0.45449901481427424,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Aquila2-34B/results_2024-04-18T14-04-47.026230.json,BAAI/Aquila2-34B,False,33693554688,0.8261661293083891,0.643049056717646,0.4471267110923455,0.4920183585480058 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Aquila2-7B/results_2024-02-13T23-17-21.976390.json,BAAI/Aquila2-7B,False,7295537152,0.5642307632772472,0.32002670516594683,0.46611451506525536,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Aquila2-7B/results_2024-04-03T05-55-31.957348.json,BAAI/Aquila2-7B,False,7295537152,0.5426094787796916,0.3589709171853071,0.49799737773227726,0.642139037433155 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Emu3-Chat/results_2024-10-23T07-24-58.039230.json,BAAI/Emu3-Chat,False,8492011520,0.3333333333333333,0.1347279162589762,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Emu3-Gen/results_2024-10-23T02-54-02.753983.json,BAAI/Emu3-Gen,False,8492011520,0.0,0.04718167761792302,0.0,0.4333858584779487 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Emu3-Stage1/results_2024-10-23T01-39-06.029347.json,BAAI/Emu3-Stage1,False,8492011520,0.5195268543314986,0.06240586658952754,0.4396551724137931,0.4946676587301587 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Gemma2-9B-IT-Simpo-Infinity-Preference/results_2024-10-04T05-31-00.057927.json,BAAI/Gemma2-9B-IT-Simpo-Infinity-Preference,False,9241705984,0.9275070649057842,0.7792788542377653,0.7059360440659401,0.8936026780915527 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Gemma2-9B-IT-Simpo-Infinity-Preference/results_2025-05-07T16-13-17.495494.json,BAAI/Gemma2-9B-IT-Simpo-Infinity-Preference,False,0,0.9308059939638886,0.7736833047344445,0.7185244587008821,0.8886125389192874 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Infinity-Instruct-3M-0613-Mistral-7B/results_2024-06-22T01-31-31.647844.json,BAAI/Infinity-Instruct-3M-0613-Mistral-7B,False,7241748480,0.9174712657490132,0.7632047672731808,0.8241841468197617,0.7990490978163615 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Infinity-Instruct-3M-0625-Llama3-8B/results_2024-07-19T01-31-48.292983.json,BAAI/Infinity-Instruct-3M-0625-Llama3-8B,False,8030261248,0.9157306525139366,0.6927579734425038,0.6831208704581523,0.8396850647140018 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Infinity-Instruct-3M-0625-Llama3-8B/results_2024-07-19T01-31-48.312064.json,BAAI/Infinity-Instruct-3M-0625-Llama3-8B,False,8030261248,0.9157306525139366,0.6927579734425038,0.6831208704581523,0.8396850647140018 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Infinity-Instruct-3M-0625-Mistral-7B/results_2024-07-19T01-41-09.242433.json,BAAI/Infinity-Instruct-3M-0625-Mistral-7B,False,7241748480,0.9195035639155449,0.7748240246646378,0.8110585067106806,0.8052263390689189 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Infinity-Instruct-3M-0625-Qwen2-7B/results_2024-07-16T18-14-59.914390.json,BAAI/Infinity-Instruct-3M-0625-Qwen2-7B,False,7615616512,0.9260247502412066,0.7493063947663584,0.7919477341597589,0.7747879604913737 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Infinity-Instruct-3M-0625-Qwen2-7B/results_2024-07-16T18-14-59.954783.json,BAAI/Infinity-Instruct-3M-0625-Qwen2-7B,False,7615616512,0.9260247502412066,0.7493063947663584,0.7919477341597589,0.7747879604913737 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Infinity-Instruct-3M-0625-Qwen2-7B/results_2024-07-16T18-15-02.013291.json,BAAI/Infinity-Instruct-3M-0625-Qwen2-7B,False,7615616512,0.9260247502412066,0.7493063947663584,0.7919477341597589,0.7747879604913737 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Infinity-Instruct-3M-0625-Yi-1.5-9B/results_2024-08-08T03-15-01.373098.json,BAAI/Infinity-Instruct-3M-0625-Yi-1.5-9B,False,8829407232,0.8898452558964334,0.7424901825788529,0.7683175374941738,0.8633798389731095 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Infinity-Instruct-3M-0625-Yi-1.5-9B/results_2025-05-04T10-35-32.472663.json,BAAI/Infinity-Instruct-3M-0625-Yi-1.5-9B,False,8829407232,0.8902591771846999,0.7417032662621263,0.7691782381060779,0.8633798389731095 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Infinity-Instruct-7M-0729-Llama3_1-8B/results_2024-08-08T04-37-11.506316.json,BAAI/Infinity-Instruct-7M-0729-Llama3_1-8B,False,8030261248,0.9353919239904989,0.7590004583613057,0.7479196445389977,0.8223021238433512 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/Infinity-Instruct-7M-Gen-Llama3_1-8B/results_2024-08-25T02-40-32.192736.json,BAAI/Infinity-Instruct-7M-Gen-Llama3_1-8B,False,8030261248,0.9353919239904989,0.7590004583613057,0.7479196445389977,0.8223021238433512 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BAAI/OPI-Llama-3.1-8B-Instruct/results_2024-10-02T03-38-58.627810.json,BAAI/OPI-Llama-3.1-8B-Instruct,False,8030261248,0.3364346429550707,0.17801095882930634,0.17721518987341772,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/baichuan-inc/Baichuan-7B/results_2024-02-13T16-00-29.790614.json,baichuan-inc/Baichuan-7B,False,7000559616,0.3333333333333333,0.16664859797501688,0.41377140792945644,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/baichuan-inc/Baichuan-7B/results_2024-04-02T22-07-44.467188.json,baichuan-inc/Baichuan-7B,False,7000559616,0.634981019204878,0.033415534259687296,0.4760391786278551,0.3688869963384406 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/baichuan-inc/Baichuan2-13B-Base/results_2024-02-13T18-20-08.619292.json,baichuan-inc/Baichuan2-13B-Base,False,13896668160,0.5360021760870939,0.1052548005298192,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/baichuan-inc/Baichuan2-13B-Base/results_2024-04-03T00-36-23.239142.json,baichuan-inc/Baichuan2-13B-Base,False,13896668160,0.7286528442256739,0.2976411689820836,0.4396551724137931,0.787803696326024 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/baichuan-inc/Baichuan2-13B-Chat/results_2024-05-27T05-34-57.596146.json,baichuan-inc/Baichuan2-13B-Chat,False,13896668160,0.7943921867307393,0.4474793720099136,0.48896075378288817,0.7634894938343515 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/baichuan-inc/Baichuan2-7B-Base/results_2024-02-13T18-20-08.619747.json,baichuan-inc/Baichuan2-7B-Base,False,7505973248,0.5418181818181818,0.18098885021409325,0.45391686964720673,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/baichuan-inc/Baichuan2-7B-Base/results_2024-04-02T23-15-16.226523.json,baichuan-inc/Baichuan2-7B-Base,False,7505973248,0.6768700922992148,0.40858850592639795,0.45391686964720673,0.4957058514404703 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/bardsai/jaskier-7b-dpo-v5.6/results_2024-02-26T22-47-54.215589.json,bardsai/jaskier-7b-dpo-v5.6,False,7241732096,0.9227847245691112,0.7770773139641775,0.7831328616487405,0.8099141363386402 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/bardsai/jaskier-7b-dpo-v5.6/results_2024-04-19T00-41-39.307543.json,bardsai/jaskier-7b-dpo-v5.6,False,7241732096,0.9223754656270444,0.776373064941006,0.7886569534429237,0.8114721156360248 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/bartowski/internlm2-chat-7b-llama/results_2024-04-25T11-11-35.087500.json,bartowski/internlm2-chat-7b-llama,False,7737708544,0.9105116047709384,0.8251651704555856,0.7855415483791437,0.8798342202712314 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/beowolx/CodeNinja-1.0-OpenChat-7B/results_2024-07-28T04-49-19.795242.json,beowolx/CodeNinja-1.0-OpenChat-7B,False,7241748480,0.9154099953342754,0.7525744119523322,0.7775191675794085,0.790679995198696 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/berkeley-nest/Starling-LM-7B-alpha/results_2024-03-08T14-01-57.709558.json,berkeley-nest/Starling-LM-7B-alpha,False,7241748480,0.927284475393028,0.8013588331495842,0.784159073392008,0.8222299935886227 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/bigscience/bloom-1b7/results_2024-02-07T01-17-14.803861.json,bigscience/bloom-1b7,False,1722408960,0.3378561857199312,0.017927392336598496,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/bigscience/bloom-1b7/results_2024-02-22T11-06-26.653643.json,bigscience/bloom-1b7,False,1722408960,0.5360021760870939,0.048115315102448805,0.4396551724137931,0.3488564106413322 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/bigscience/bloom-3b/results_2024-02-07T01-42-50.987972.json,bigscience/bloom-3b,False,3002557440,0.3333333333333333,0.007784203462752267,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/bigscience/bloom-3b/results_2024-03-08T11-23-13.127097.json,bigscience/bloom-3b,False,3002557440,0.5105618421880478,0.06267758607338429,0.4396551724137931,0.6118897273987371 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/bigscience/bloom-560m/results_2024-02-07T01-17-10.201449.json,bigscience/bloom-560m,False,559214592,0.3333333333333333,0.057306527654191086,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/bigscience/bloom-560m/results_2024-02-25T19-07-46.918697.json,bigscience/bloom-560m,False,559214592,0.3333333333333333,0.08478982085671935,0.4396551724137931,0.37073579075474156 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/bigscience/bloom-7b1/results_2024-02-07T01-49-37.829344.json,bigscience/bloom-7b1,False,7069016064,0.43542766255781346,0.09832383754488921,0.47143349150466857,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/bigscience/bloom-7b1/results_2024-02-22T11-33-09.451361.json,bigscience/bloom-7b1,False,7069016064,0.3483892974147847,0.0940150443614523,0.4518479958355024,0.4930101744884911 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BornSaint/Dare_Angel_8B/results_2025-07-07T06-21-20.661894.json,BornSaint/Dare_Angel_8B,False,8030261248,0.919089077733271,0.7264808096329166,0.7446341747047971,0.8198582099698948 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/botbot-ai/Cabra-72b/results_2024-05-19T21-21-24.695587.json,botbot-ai/Cabra-72b,False,72281071616,0.9358480179428454,0.7803353287613494,0.45449901481427424,0.7211537359669663 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/botbot-ai/CabraLlama3-70b/results_2024-07-07T20-18-29.780323.json,botbot-ai/CabraLlama3-70b,False,70553706496,0.9321018018499706,0.8332401038565972,0.8059707857281098,0.8162058158781371 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/botbot-ai/CabraLlama3-8b/results_2024-04-23T06-06-54.135257.json,botbot-ai/CabraLlama3-8b,False,8030261248,0.9043933220596747,0.6985414491675039,0.7038187988033884,0.850539126721469 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/botbot-ai/CabraMistral-v3-7b-32k/results_2024-06-11T23-54-09.171137.json,botbot-ai/CabraMistral-v3-7b-32k,False,7248023552,0.8642982568523864,0.6833344577403617,0.4682788475891924,0.7109141849844915 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/botbot-ai/CabraMistral-v3-7b-32k/results_2024-06-12T05-18-28.387029.json,botbot-ai/CabraMistral-v3-7b-32k,False,7248023552,0.861379516929838,0.6806065734433028,0.47460337127045094,0.7045761256287573 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/botbot-ai/CabraMixtral-8x7b/results_2024-05-19T01-59-57.504600.json,botbot-ai/CabraMixtral-8x7b,False,46702792704,0.9095362286997337,0.7763068289471359,0.7893132508517124,0.7799738064055245 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Bruno/Caramelinho/results_2024-02-26T15-17-54.708968.json,Bruno/Caramelinho,False,7052268416,0.4896626375608876,0.19384903999896694,0.43917169974115616,0.3396512838306731 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Bruno/Caramelo_7B/results_2024-02-26T13-57-57.036659.json,Bruno/Caramelo_7B,False,7052268416,0.5427381481762671,0.07473225338478715,0.4396551724137931,0.33650009913117634 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BSC-LT/salamandra-2b/results_2025-04-06T03-29-18.305527.json,BSC-LT/salamandra-2b,False,2253490176,0.6975027707196586,0.0671816933521037,0.4396551724137931,0.5042049852088193 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BSC-LT/salamandra-7b/results_2025-04-06T04-40-14.780421.json,BSC-LT/salamandra-7b,False,7768117248,0.5473539470436837,0.37404739198943987,0.5425562583730434,0.7970481703639771 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/BSC-LT/salamandra-7b-instruct/results_2025-04-07T04-07-55.633968.json,BSC-LT/salamandra-7b-instruct,False,7768117248,0.9098606334526578,0.6729342980455344,0.5734435078379178,0.8052173913043479 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/bunnycore/HyperLlama-3.1-8B/results_2025-02-04T20-10-07.941710.json,bunnycore/HyperLlama-3.1-8B,False,8030261248,0.9214671443083997,0.7751168118375961,0.74,0.8566017482295437 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/byroneverson/gemma-2-27b-it-abliterated/results_2024-10-20T02-43-14.061434.json,byroneverson/gemma-2-27b-it-abliterated,False,27227128320,0.9130271568921875,0.7769289884131394,0.7650872473963721,0.8792816023519557 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/byroneverson/gemma-2-27b-it-abliterated/results_2025-07-07T10-49-34.604310.json,byroneverson/gemma-2-27b-it-abliterated,False,27227128320,0.9138568369265416,0.7489499257157044,0.7680440176933865,0.8792816023519557 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/byroneverson/Mistral-Small-Instruct-2409-abliterated/results_2024-10-14T03-44-56.687411.json,byroneverson/Mistral-Small-Instruct-2409-abliterated,False,22247282688,0.9341369689951045,0.7908328744051527,0.8607793063679051,0.8414878124279372 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/byroneverson/Yi-1.5-9B-Chat-16K-abliterated/results_2024-10-06T02-38-37.849161.json,byroneverson/Yi-1.5-9B-Chat-16K-abliterated,False,8829407232,0.8872512143496513,0.7064923614014677,0.6764069264069263,0.8416301863402162 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/byroneverson/Yi-1.5-9B-Chat-abliterated/results_2024-10-18T03-43-36.257686.json,byroneverson/Yi-1.5-9B-Chat-abliterated,False,8829407232,0.8798248748509733,0.7412745664075343,0.7529442144264331,0.7841024224002947 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CausalLM/34b-beta/results_2024-06-13T11-57-13.556700.json,CausalLM/34b-beta,False,34388917248,0.876489392618425,0.772190146473889,0.751649303344456,0.8089822265646441 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CEIA-UFG/Gemma-3-Gaia-PT-BR-4b-it/results_2025-07-08T04-44-42.139045.json,CEIA-UFG/Gemma-3-Gaia-PT-BR-4b-it,False,4300079472,0.9044663406081531,0.728790961400952,0.6252696875023602,0.8805545331861121 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cerebras/btlm-3b-8k-base/results_2024-04-24T18-05-47.789710.json,cerebras/btlm-3b-8k-base,False,2646255776,0.6560035708971879,0.2347691630838717,0.5528476363322098,0.5403327115611558 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cerebras/Cerebras-GPT-1.3B/results_2024-06-13T10-23-29.287780.json,cerebras/Cerebras-GPT-1.3B,False,1315723264,0.33424046164366045,0.05363861584121788,0.4396551724137931,0.4630444643316865 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cerebras/Cerebras-GPT-111M/results_2024-06-13T08-16-48.589605.json,cerebras/Cerebras-GPT-111M,False,111050496,0.3333333333333333,0.0,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cerebras/Cerebras-GPT-13B/results_2024-06-30T01-31-27.078874.json,cerebras/Cerebras-GPT-13B,False,12853386240,0.3369541665038594,0.14736179615891565,0.30803682567080237,0.3389990557129367 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cerebras/Cerebras-GPT-2.7B/results_2024-06-13T11-32-13.549041.json,cerebras/Cerebras-GPT-2.7B,False,2651553280,0.33424046164366045,0.09492307926905007,0.4396551724137931,0.34205234732031575 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cerebras/Cerebras-GPT-256M/results_2024-06-13T08-50-29.374793.json,cerebras/Cerebras-GPT-256M,False,255977024,0.33424046164366045,0.0,0.4396551724137931,0.49124536178107603 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cerebras/Cerebras-GPT-590M/results_2024-06-13T09-31-31.581772.json,cerebras/Cerebras-GPT-590M,False,590310912,0.3333333333333333,0.0131636700526589,0.4396551724137931,0.34212658391936995 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cerebras/Cerebras-GPT-6.7B/results_2024-06-13T11-35-08.467894.json,cerebras/Cerebras-GPT-6.7B,False,6658404352,0.3851604537249648,0.1560580681300574,0.17721518987341772,0.38621441861459826 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/chargoddard/internlm2-base-7b-llama/results_2024-04-25T07-01-22.107597.json,chargoddard/internlm2-base-7b-llama,False,7737708544,0.7328476894845132,0.6780142465261234,0.4396551724137931,0.5150166743512616 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/chujiezheng/Llama-3-Instruct-8B-SimPO-ExPO/results_2024-05-30T03-47-51.756674.json,chujiezheng/Llama-3-Instruct-8B-SimPO-ExPO,False,8030261248,0.8989008703317252,0.7162092421813775,0.5873015873015873,0.8148684191345671 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/chujiezheng/Mistral7B-PairRM-SPPO-ExPO/results_2024-10-02T05-58-49.373970.json,chujiezheng/Mistral7B-PairRM-SPPO-ExPO,False,7241732096,0.8920213313957587,0.695411691341475,0.5892857142857143,0.7369482256668503 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/chujiezheng/Smaug-34B-v0.1-ExPO/results_2024-06-13T19-09-40.173329.json,chujiezheng/Smaug-34B-v0.1-ExPO,False,34388917248,0.9158370333876075,0.7758033092113542,0.7538654421989506,0.6461593212424546 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cnmoro/Mistral-7B-Portuguese/results_2024-03-08T01-01-24.000401.json,cnmoro/Mistral-7B-Portuguese,False,7241732096,0.9031337575789451,0.7654589376771873,0.5884431791136597,0.7921019662468967 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cnmoro/Mistral-7B-Portuguese/results_2024-04-20T04-17-11.730188.json,cnmoro/Mistral-7B-Portuguese,False,7241732096,0.9035406238416921,0.7630986536475501,0.5975052901213944,0.7907328739630719 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cnmoro/Qwen2.5-0.5B-Portuguese-Hybrid-Reasoning/results_2025-04-03T21-58-07.753268.json,cnmoro/Qwen2.5-0.5B-Portuguese-Hybrid-Reasoning,False,494032768,0.8267494803825359,0.4724530683545055,0.6331981774861012,0.36542668169365256 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cnmoro/Qwen2.5-0.5B-Portuguese-v1/results_2025-04-03T20-56-24.174479.json,cnmoro/Qwen2.5-0.5B-Portuguese-v1,False,494032768,0.8630287011918854,0.5429592652527777,0.6533009410124937,0.44060362771335115 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cnmoro/Qwen2.5-0.5B-Portuguese-v2/results_2025-04-03T21-20-29.740963.json,cnmoro/Qwen2.5-0.5B-Portuguese-v2,False,494032768,0.8790768959652246,0.590104418454831,0.4396551724137931,0.3361782835467046 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/dolphin-2.9-llama3-8b/results_2024-04-22T02-16-15.555305.json,cognitivecomputations/dolphin-2.9-llama3-8b,False,8030277632,0.9006039149623303,0.7810623480752894,0.724312599217169,0.7732619591530567 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/dolphin-2.9.1-llama-3-70b/results_2024-07-08T01-03-47.577530.json,cognitivecomputations/dolphin-2.9.1-llama-3-70b,False,70553739264,0.9211474857983728,0.7825824088338631,0.5275390900126778,0.8100949263502455 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/dolphin-2.9.1-mixtral-1x22b/results_2024-05-26T22-36-26.076147.json,cognitivecomputations/dolphin-2.9.1-mixtral-1x22b,False,22238214144,0.8688714768328896,0.6796168190100729,0.5948262427925822,0.6872950365323062 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/dolphin-2.9.1-yi-1.5-34b/results_2024-05-25T11-37-37.984087.json,cognitivecomputations/dolphin-2.9.1-yi-1.5-34b,False,34388917248,0.911738725475468,0.825224950259231,0.7980769230769231,0.8468375373508494 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/dolphin-2.9.1-yi-1.5-9b/results_2024-05-25T12-44-50.874130.json,cognitivecomputations/dolphin-2.9.1-yi-1.5-9b,False,8829407232,0.906826923076923,0.7593694153525102,0.6903352848386037,0.8696157656440613 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/dolphin-2.9.2-Phi-3-Medium/results_2024-06-18T13-12-24.761209.json,cognitivecomputations/dolphin-2.9.2-Phi-3-Medium,False,13960238080,0.9358657455204926,0.726807568078287,0.7670573030568968,0.8009985869846759 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/dolphin-2.9.2-Phi-3-Medium-abliterated/results_2024-06-17T07-33-05.284826.json,cognitivecomputations/dolphin-2.9.2-Phi-3-Medium-abliterated,False,13960238080,0.9362702559816725,0.7297023607932795,0.7814473045167557,0.7250569715248458 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/dolphin-2.9.2-qwen2-7b/results_2024-06-18T13-11-30.847393.json,cognitivecomputations/dolphin-2.9.2-qwen2-7b,False,7615616512,0.9334069317712441,0.7617869377416114,0.6575532386867791,0.6665618718263835 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/dolphin-2.9.3-llama-3-8b/results_2024-07-03T01-36-26.198500.json,cognitivecomputations/dolphin-2.9.3-llama-3-8b,False,8030277632,0.8458152184410374,0.7122883766333097,0.6050568832223867,0.5917288642225078 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/dolphin-2.9.3-mistral-7B-32k/results_2024-07-01T01-31-45.295693.json,cognitivecomputations/dolphin-2.9.3-mistral-7B-32k,False,7248039936,0.9072711025673759,0.7700090320870142,0.740456116464181,0.8124860915024096 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/dolphin-2.9.3-mistral-7B-32k/results_2024-07-01T01-31-45.855293.json,cognitivecomputations/dolphin-2.9.3-mistral-7B-32k,False,7248039936,0.9072711025673759,0.7700090320870142,0.740456116464181,0.8124860915024096 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/dolphin-2.9.3-mistral-nemo-12b/results_2024-07-26T04-21-48.617468.json,cognitivecomputations/dolphin-2.9.3-mistral-nemo-12b,False,12247802880,0.9308416198672873,0.8082965505115898,0.8128047508863223,0.8585034013605441 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/dolphin-2.9.4-llama3.1-8b/results_2024-08-26T01-39-29.053763.json,cognitivecomputations/dolphin-2.9.4-llama3.1-8b,False,8030277632,0.6827646558194831,0.11371065302601087,0.4396551724137931,0.7048751879044493 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/Dolphin3.0-R1-Mistral-24B/results_2025-02-12T17-49-30.090166.json,cognitivecomputations/Dolphin3.0-R1-Mistral-24B,False,23572423680,0.9018206257032176,0.7681507891980223,0.66761929235745,0.8843243004673729 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/laserxtral/results_2024-08-12T01-36-06.572396.json,cognitivecomputations/laserxtral,False,24153690112,0.9231594800094836,0.7829095943476577,0.766772890921475,0.8362695115615253 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/openchat-3.5-0106-laser/results_2024-08-08T01-36-00.321142.json,cognitivecomputations/openchat-3.5-0106-laser,False,7241748480,0.9248076923076923,0.8345948997950121,0.7866754184443715,0.7732668320172721 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/WestLake-7B-v2-laser/results_2024-08-08T01-40-30.831958.json,cognitivecomputations/WestLake-7B-v2-laser,False,7241732096,0.9190722183078603,0.7771685824919177,0.7046043452262308,0.8711919852790839 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/WizardLM-13B-Uncensored/results_2024-05-22T01-59-20.847149.json,cognitivecomputations/WizardLM-13B-Uncensored,False,13015874560,0.6660477603775113,0.3828951397360637,0.5550602028995683,0.7522345118834984 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/WizardLM-30B-Uncensored/results_2024-08-12T01-05-43.317704.json,cognitivecomputations/WizardLM-30B-Uncensored,False,32528956928,0.7886430698387075,0.7060765182168839,0.65606529693217,0.7633469292424382 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivecomputations/WizardLM-7B-Uncensored/results_2024-05-22T03-30-08.099090.json,cognitivecomputations/WizardLM-7B-Uncensored,False,6738423808,0.6793655049395517,0.17381621240651648,0.5274657953849295,0.7837781996151809 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cognitivess/bella-1-8b/results_2024-06-29T02-19-40.028800.json,cognitivess/bella-1-8b,False,8072204288,0.9374443725338734,0.7701455647086599,0.7816849545960171,0.8448552638392626 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CohereForAI/aya-101/results_2024-02-21T19-25-38.847154.json,CohereForAI/aya-101,False,12921057280,0.845896116707975,0.18932506997017534,0.3536861536119358,0.8577866430260047 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CohereForAI/aya-23-35B/results_2024-05-25T09-17-37.504596.json,CohereForAI/aya-23-35B,False,34980831232,0.9304615643741841,0.7846161558721925,0.7233650163143708,0.8860471199766156 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CohereForAI/aya-23-8B/results_2024-05-25T07-14-27.654611.json,CohereForAI/aya-23-8B,False,8028033024,0.9189769820971867,0.780672309349922,0.6541835357624831,0.7471163522824039 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CohereForAI/aya-expanse-32b/results_2024-12-03T17-11-02.917534.json,CohereForAI/aya-expanse-32b,False,32296476672,0.923384962213484,0.7373015447324608,0.720673391991921,0.8251623100939243 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CohereForAI/aya-expanse-8b/results_2024-12-04T11-05-27.239366.json,CohereForAI/aya-expanse-8b,False,8028033024,0.9115861876158329,0.7797645180627574,0.7434885556432518,0.8207730793849852 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CohereForAI/c4ai-command-r-plus-4bit/results_2024-04-14T07-15-54.067150.json,CohereForAI/c4ai-command-r-plus-4bit,False,103809822720,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CohereForAI/c4ai-command-r-plus-4bit/results_2024-04-15T16-05-38.445928.json,CohereForAI/c4ai-command-r-plus-4bit,False,103810674688,0.9301234467745643,0.7933785386356376,0.7718257450767017,0.773798484417851 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CohereForAI/c4ai-command-r-v01/results_2024-04-17T00-36-42.568466.json,CohereForAI/c4ai-command-r-v01,False,34980831232,0.883132179380006,0.7210331309303998,0.47272296015180265,0.8222299935886227 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Columbia-NLP/LION-LLaMA-3-8b-odpo-v1.0/results_2024-07-15T01-32-05.828202.json,Columbia-NLP/LION-LLaMA-3-8b-odpo-v1.0,False,8030261248,0.9193834267092047,0.7172104868084787,0.7286917112711735,0.8564797460211679 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CombinHorizon/YiSM-blossom5.1-34B-SLERP/results_2024-12-08T13-52-36.707939.json,CombinHorizon/YiSM-blossom5.1-34B-SLERP,False,34388917248,0.9215516619799553,0.8275533099042904,0.8305637015314435,0.881900537208532 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ConvexAI/Luminex-34B-v0.1/results_2024-06-14T16-37-48.304200.json,ConvexAI/Luminex-34B-v0.1,False,34388917248,0.9191070641797621,0.8130683879495547,0.8226956044555551,0.6983754481802518 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ConvexAI/Luminex-34B-v0.2/results_2024-06-13T15-37-38.009373.json,ConvexAI/Luminex-34B-v0.2,False,34388917248,0.9162531250297254,0.8034116288676774,0.8054074649748557,0.6992729676140119 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/cosmicvalor/mistral-orthogonalized/results_2024-05-25T16-02-13.102801.json,cosmicvalor/mistral-orthogonalized,False,7241732096,0.9023493197589443,0.7647667587328075,0.7445364698077083,0.8245222600326303 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Cran-May/T.E-8.1/results_2024-10-05T04-54-18.066733.json,Cran-May/T.E-8.1,False,7615616512,0.9342280640477356,0.7928489762220716,0.7432359307359307,0.7896677858027337 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/croissantllm/CroissantLLMBase/results_2024-02-24T19-45-27.123677.json,croissantllm/CroissantLLMBase,False,1345423360,0.3333333333333333,0.13488403072117375,0.4396551724137931,0.3349186726374015 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CultriX/NeuralMona_MoE-4x7B/results_2024-05-26T13-29-26.736769.json,CultriX/NeuralMona_MoE-4x7B,False,24153690112,0.9244279910791389,0.7720274719342004,0.7694314032342202,0.8409826856991804 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CultriX/Qwen2.5-14B-Wernicke/results_2024-11-19T02-46-31.548628.json,CultriX/Qwen2.5-14B-Wernicke,False,14770033664,0.9444324272358017,0.8370321624309986,0.7906626549720158,0.890936897154724 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CultriX/Qwen2.5-14B-Wernicke-SFT/results_2024-11-19T04-37-24.664140.json,CultriX/Qwen2.5-14B-Wernicke-SFT,False,14770033664,0.9436210912232053,0.8319345766276613,0.8653846153846154,0.802972634621363 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/CultriX/SeQwence-14Bv1/results_2024-12-21T21-30-28.324802.json,CultriX/SeQwence-14Bv1,False,14765947904,0.9456698530460836,0.7903813479269409,0.7833745953151781,0.9055290068890212 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/DAMO-NLP-MT/polylm-1.7b/results_2024-02-16T14-47-24.223296.json,DAMO-NLP-MT/polylm-1.7b,False,1737084928,0.3333333333333333,0.004714541859212106,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/DAMO-NLP-MT/polylm-1.7b/results_2024-04-17T23-46-04.491918.json,DAMO-NLP-MT/polylm-1.7b,False,1737084928,0.4047692251758633,0.05167868234986358,0.4396551724137931,0.358843537414966 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/DAMO-NLP-MT/polylm-13b/results_2024-04-03T09-53-29.935717.json,DAMO-NLP-MT/polylm-13b,False,13906790400,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Dampfinchen/Llama-3.1-8B-Ultra-Instruct/results_2024-09-20T04-18-17.935178.json,Dampfinchen/Llama-3.1-8B-Ultra-Instruct,False,8030261248,0.9235696737015304,0.7286168616739841,0.7912633013135291,0.8682961944962534 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Danielbrdz/Barcenas-14b-Phi-3-medium-ORPO/results_2024-08-08T02-43-35.640819.json,Danielbrdz/Barcenas-14b-Phi-3-medium-ORPO,False,13960238080,0.9259642329554806,0.714480317302389,0.6906170752324599,0.8460180802244769 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Danielbrdz/Barcenas-Llama3-8b-ORPO/results_2024-05-18T00-12-52.690138.json,Danielbrdz/Barcenas-Llama3-8b-ORPO,False,8030261248,0.9178150146340144,0.7260402501200387,0.7308849598805747,0.8698828946051447 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Deci/DeciLM-7B/results_2024-02-07T06-30-45.967063.json,Deci/DeciLM-7B,False,7043551232,0.7470003950561543,0.5145659724171266,0.7851487912814374,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Deci/DeciLM-7B/results_2024-04-02T05-42-17.715000.json,Deci/DeciLM-7B,False,7043551232,0.9123267863598024,0.7555893659678592,0.7857378310075815,0.6990533471973728 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepcogito/cogito-v1-preview-llama-3B/results_2025-04-10T05-11-54.217825.json,deepcogito/cogito-v1-preview-llama-3B,False,3212749824,0.9038873914685859,0.7026305829345696,0.6918229995146599,0.8439225806451613 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepcogito/cogito-v1-preview-llama-70B/results_2025-07-15T07-04-56.994849.json,deepcogito/cogito-v1-preview-llama-70B,False,70553706496,0.9317329158205281,0.7305423388399148,0.8274180003951788,0.8796606409486324 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepcogito/cogito-v1-preview-llama-8B/results_2025-04-11T06-29-02.373217.json,deepcogito/cogito-v1-preview-llama-8B,False,8030261248,0.9313708999158957,0.7251698884137406,0.7642199365792584,0.8869488244108766 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/DeepMount00/Llama-3-8b-Ita/results_2024-05-19T23-04-56.757278.json,DeepMount00/Llama-3-8b-Ita,False,8030261248,0.9174373136272109,0.7274385305362817,0.7530594601709288,0.8684530363041023 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/DeepMount00/Llama-3.1-8b-Ita/results_2024-09-22T07-13-23.970075.json,DeepMount00/Llama-3.1-8b-Ita,False,8030261248,0.9231533213332754,0.7337286719556836,0.7017672651113582,0.8705424903149537 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/deepseek-llm-67b-base/results_2024-02-12T12-58-00.419466.json,deepseek-ai/deepseek-llm-67b-base,False,67425001472,0.6950181531654831,0.5337084187331966,0.6754332808144468,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/deepseek-llm-7b-base/results_2024-02-13T20-08-54.493662.json,deepseek-ai/deepseek-llm-7b-base,False,6910365696,0.36087257670772555,0.3246306803889829,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/deepseek-llm-7b-base/results_2024-04-03T04-04-12.490520.json,deepseek-ai/deepseek-llm-7b-base,False,6910365696,0.4641752126120648,0.23484125545266332,0.4396551724137931,0.5198841409416219 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/deepseek-moe-16b-base/results_2024-02-13T21-07-50.712249.json,deepseek-ai/deepseek-moe-16b-base,False,16375728128,0.3669298673297707,0.2023483280938947,0.45391686964720673,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/deepseek-moe-16b-base/results_2024-04-18T16-50-06.702500.json,deepseek-ai/deepseek-moe-16b-base,False,16375728128,0.7622492283950617,0.17844225740007624,0.4471267110923455,0.6096281194190255 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/deepseek-moe-16b-chat/results_2024-04-19T08-15-02.021043.json,deepseek-ai/deepseek-moe-16b-chat,False,16375728128,0.7768240314630341,0.5333994008721673,0.6171743795484232,0.6510610271031073 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/results_2025-02-03T13-40-03.313124.json,deepseek-ai/DeepSeek-R1-Distill-Llama-70B,False,0,0.8300781128111251,0.5112130961735681,0.4789067296779355,0.5576926076858632 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/results_2025-05-07T00-50-36.869893.json,deepseek-ai/DeepSeek-R1-Distill-Llama-70B,False,0,0.7775908887605295,0.42069323790791385,0.5268647889406373,0.6599913203934313 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Llama-8B/results_2025-01-29T08-57-22.682676.json,deepseek-ai/DeepSeek-R1-Distill-Llama-8B,False,8030261248,0.6250476257260732,0.5294190448361772,0.0019569471624266144,0.8391114119012433 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Llama-8B/results_2025-01-30T01-50-42.314640.json,deepseek-ai/DeepSeek-R1-Distill-Llama-8B,False,8030261248,0.6250476257260732,0.0,0.0313925442707357,0.514634878006971 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Llama-8B/results_2025-02-03T01-17-27.055431.json,deepseek-ai/DeepSeek-R1-Distill-Llama-8B,False,0,0.6644900111797395,0.5569868073960826,0.6751684521037106,0.7767453588974392 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Llama-8B/results_2025-05-06T03-31-14.191270.json,deepseek-ai/DeepSeek-R1-Distill-Llama-8B,False,0,0.7008676005568326,0.29058664477179313,0.6661966963853756,0.6790568067741369 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B/results_2025-01-29T05-01-04.680013.json,deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B,False,1777088000,0.3260741696336983,0.06696072170398638,0.40719696969696967,0.6339089457037985 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B/results_2025-01-29T20-33-41.335461.json,deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B,False,1777088000,0.3675065923701736,0.004714541859211954,0.3905415713196034,0.6339089457037985 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B/results_2025-02-03T12-25-24.538017.json,deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B,False,0,0.5268715747196657,0.04717763988064833,0.37884926008949105,0.5670041642312024 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B/results_2025-05-05T18-40-12.480025.json,deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B,False,0,0.4773265359025317,0.08094318232305112,0.26363636363636367,0.44509754319132155 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B/results_2025-02-03T03-06-52.213625.json,deepseek-ai/DeepSeek-R1-Distill-Qwen-14B,False,0,0.8543605075812231,0.720360932470465,0.7277238089168495,0.8778077082469955 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B/results_2025-05-06T13-58-33.053796.json,deepseek-ai/DeepSeek-R1-Distill-Qwen-14B,False,0,0.8383857496718465,0.5659012454050178,0.7177502249860317,0.6615789504881244 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B/results_2025-01-29T11-04-30.630214.json,deepseek-ai/DeepSeek-R1-Distill-Qwen-32B,False,32763876352,0.9166780541646973,0.7305117471973348,0.7527225155249755,0.9270772266468393 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B/results_2025-01-30T07-03-24.143472.json,deepseek-ai/DeepSeek-R1-Distill-Qwen-32B,False,32763876352,0.9166780541646973,0.7305117471973348,0.7527225155249755,0.9270772266468393 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B/results_2025-02-02T09-26-56.899617.json,deepseek-ai/DeepSeek-R1-Distill-Qwen-32B,False,0,0.9174800323674799,0.6799738702298108,0.7068073305314921,0.9263229802303403 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B/results_2025-01-29T06-45-53.340493.json,deepseek-ai/DeepSeek-R1-Distill-Qwen-7B,False,7615616512,0.08588498879761015,0.529314859689523,0.2968152866242038,0.8028567405239603 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B/results_2025-02-02T23-53-07.620072.json,deepseek-ai/DeepSeek-R1-Distill-Qwen-7B,False,0,0.771308337573567,0.5058163344527936,0.4662875851776438,0.47535907155512125 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B/results_2025-05-06T02-35-11.882597.json,deepseek-ai/DeepSeek-R1-Distill-Qwen-7B,False,0,0.7076872039347348,0.34483124264228193,0.5071176518542386,0.582749871923105 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-V2-Lite/results_2024-05-18T00-55-21.183568.json,deepseek-ai/DeepSeek-V2-Lite,False,15706484224,0.7433517408575088,0.28354183355746143,0.4396551724137931,0.48732033324178337 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/deepseek-ai/DeepSeek-V2-Lite-Chat/results_2024-05-18T01-13-35.002008.json,deepseek-ai/DeepSeek-V2-Lite-Chat,False,15706484224,0.8741376581682254,0.6776897775296321,0.712782196513216,0.7091563407352881 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Delta-Vector/Odin-9B/results_2024-11-02T01-41-08.746459.json,Delta-Vector/Odin-9B,False,9241705984,0.42999722804440965,0.005347618883489636,0.4396551724137931,0.47852043543078954 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/dfurman/Llama-3-8B-Orpo-v0.1/results_2024-08-25T03-50-45.916779.json,dfurman/Llama-3-8B-Orpo-v0.1,False,8030277632,0.8215472466057163,0.30907900715064496,0.4396551724137931,0.34401702477017637 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/dnhkng/RYS-Llama-3.1-8B-Instruct/results_2024-08-25T05-52-00.600465.json,dnhkng/RYS-Llama-3.1-8B-Instruct,False,8684597248,0.9276859300688284,0.7696386743993324,0.7601389854973075,0.8562876218475663 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/dominguesm/canarim-7b/results_2024-02-06T14-23-58.409877.json,dominguesm/canarim-7b,False,6738415616,0.381404336182431,0.20436567575697118,0.48515878307149657,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/dominguesm/canarim-7b/results_2024-02-16T23-16-15.447588.json,dominguesm/canarim-7b,False,6738415616,0.719628678839795,0.13339286550512713,0.4909142132694746,0.7848804769588817 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/dominguesm/Canarim-7B-Instruct/results_2024-02-25T12-22-01.120661.json,dominguesm/Canarim-7B-Instruct,False,6738415616,0.7573936891724624,0.12077328686163717,0.43917169974115616,0.7957259180088587 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/dominguesm/mambarim-110m/results_2024-04-18T22-38-53.882980.json,dominguesm/mambarim-110m,False,69836544,0.24134985049124305,0.018889657528727573,0.1393003180372558,0.2362012987012987 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/dynamofl/dynamo-8B-v0.1/results_2024-02-06T22-12-03.723530.json,dynamofl/dynamo-8B-v0.1,False,8272482304,0.3333333333333333,0.16943917624352078,0.17721518987341772,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/dynamofl/dynamo-8B-v0.1/results_2024-02-17T03-42-08.504769.json,dynamofl/dynamo-8B-v0.1,False,8272482304,0.3333333333333333,0.2240463818083142,0.17721518987341772,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/dzakwan/dzakwan-MoE-4x7b-Beta/results_2024-06-13T02-34-28.888859.json,dzakwan/dzakwan-MoE-4x7b-Beta,False,24153690112,0.6254985943049329,0.5728700306043661,0.5516489252623926,0.6730588787239034 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/gemma-7b-it_no_chat_template/results_2024-02-23T22-19-59.816101.json,eduagarcia/gemma-7b-it_no_chat_template,False,8537680896,0.8471090065645221,0.6824898498454515,0.5713532115853199,0.8272202378666258 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/gemma-7b-it_singleturn_chat_template/results_2024-02-24T16-55-10.979216.json,eduagarcia/gemma-7b-it_singleturn_chat_template,False,8537680896,0.3333333333333333,0.0,0.17721518987341772,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/mistral-orpo-capybara-3k/results_2024-05-01T04-50-38.819939.json,eduagarcia/mistral-orpo-capybara-3k,False,7241732096,0.8664165531493728,0.7094171991678102,0.7259902741602806,0.7549884010466452 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/mistral-orpo-mix-21k/results_2024-05-06T02-46-03.110290.json,eduagarcia/mistral-orpo-mix-21k,False,7241732096,0.8889986321597236,0.7109901510044454,0.6913632586613305,0.81439532334673 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/mistral-orpo-mix-7k/results_2024-04-29T19-28-31.425145.json,eduagarcia/mistral-orpo-mix-7k,False,7241732096,0.8794910308292768,0.6280077150982621,0.6588300266089758,0.733495112908662 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/zephyr-pt-llama-3-8B-orpo-v0.1/results_2024-07-09T23-00-31.743266.json,eduagarcia/zephyr-pt-llama-3-8B-orpo-v0.1,False,8030261248,0.3720762801762666,0.23408865429807973,0.4396551724137931,0.48183332305224846 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/zephyr-pt-phi-3-mini-128k-dpo-v0.1/results_2024-04-29T04-41-26.501612.json,eduagarcia/zephyr-pt-phi-3-mini-128k-dpo-v0.1,False,3821079552,0.8990426637979585,0.7365102577010455,0.6746261926870183,0.7838776522010813 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/zephyr-pt-phi-3-mini-128k-orpo-v0.1/results_2024-04-29T11-26-31.873593.json,eduagarcia/zephyr-pt-phi-3-mini-128k-orpo-v0.1,False,3821079552,0.9015095399862072,0.7313341965089064,0.5314377252703244,0.7894649696013196 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/zephyr-pt-phi-3-mini-128k-sft-v0.1/results_2024-04-28T21-55-33.977793.json,eduagarcia/zephyr-pt-phi-3-mini-128k-sft-v0.1,False,3821079552,0.8965818323903127,0.7349182752797895,0.6711333251563841,0.783990550771436 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/zephyr-pt-Yi-1.5-9B-orpo-v0.1/results_2024-07-10T01-13-21.782307.json,eduagarcia/zephyr-pt-Yi-1.5-9B-orpo-v0.1,False,8829431808,0.8917013613956221,0.7421126526474375,0.6298334008995965,0.8350418286791563 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/zephyr-pt-Yi-1.5-9B-orpo-v0.2/results_2024-07-12T02-18-06.576935.json,eduagarcia/zephyr-pt-Yi-1.5-9B-orpo-v0.2,False,8829431808,0.8888674512830408,0.7372829956305684,0.6513636363636364,0.848963085345104 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/zephyr-pt-Yi-1.5-9B-orpo-v0.2/results_2024-07-12T02-18-12.615096.json,eduagarcia/zephyr-pt-Yi-1.5-9B-orpo-v0.2,False,8829431808,0.8888674512830408,0.7372829956305684,0.6513636363636364,0.848963085345104 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/zephyr-pt-Yi-1.5-9B-orpo-v0.3/results_2024-07-12T03-50-51.541552.json,eduagarcia/zephyr-pt-Yi-1.5-9B-orpo-v0.3,False,8829431808,0.8851625756685653,0.7375410609265864,0.6544391281233386,0.835948707386283 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/zephyr-pt-Yi-1.5-9B-orpo-v0.4/results_2024-07-12T04-15-25.317416.json,eduagarcia/zephyr-pt-Yi-1.5-9B-orpo-v0.4,False,8829431808,0.8855805139662417,0.7362738685932091,0.6690909090909091,0.8361956148050713 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/zephyr-pt-Yi-1.5-9B-orpo-v0.5/results_2024-07-12T05-12-30.921802.json,eduagarcia/zephyr-pt-Yi-1.5-9B-orpo-v0.5,False,8829431808,0.864838004372888,0.7076460361967238,0.6268942292974131,0.7845586297760211 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eduagarcia/zephyr-pt-Yi-1.5-9B-orpo-v0.6/results_2024-07-12T05-56-39.688142.json,eduagarcia/zephyr-pt-Yi-1.5-9B-orpo-v0.6,False,8829431808,0.8863502157619805,0.7769602352610964,0.661962866368604,0.7462551308705154 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/eldogbbhed/Peagle-9b/results_2024-05-19T21-23-39.265807.json,eldogbbhed/Peagle-9b,False,8986628096,0.5784120668038546,0.5522260924205177,0.5708602851459994,0.5393188795961035 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/gpt-j-6b/results_2024-02-07T15-56-06.233510.json,EleutherAI/gpt-j-6b,False,6050882784,0.3512359875045344,0.09337260889279221,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/gpt-j-6b/results_2024-04-05T04-41-08.855450.json,EleutherAI/gpt-j-6b,False,6050882784,0.3582588385476761,0.14562487212003206,0.4396551724137931,0.6588376162844248 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/gpt-neo-1.3B/results_2024-02-14T20-42-35.885763.json,EleutherAI/gpt-neo-1.3B,False,1315575808,0.3333333333333333,0.0070782273697892,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/gpt-neo-1.3B/results_2024-04-04T01-04-14.137713.json,EleutherAI/gpt-neo-1.3B,False,1315575808,0.3333333333333333,0.27954490493177114,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/gpt-neo-125m/results_2024-02-14T20-18-41.235422.json,EleutherAI/gpt-neo-125m,False,125198592,0.3333333333333333,0.10713084465982439,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/gpt-neo-125m/results_2024-04-04T00-23-23.313643.json,EleutherAI/gpt-neo-125m,False,125198592,0.40826127460837947,0.13567407692821803,0.4396551724137931,0.3359391417643845 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/gpt-neo-2.7B/results_2024-02-14T21-23-17.848186.json,EleutherAI/gpt-neo-2.7B,False,2651307520,0.3333333333333333,0.006388450736995319,0.42535491344201753,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/gpt-neo-2.7B/results_2024-04-04T01-08-46.345259.json,EleutherAI/gpt-neo-2.7B,False,2651307520,0.34680711177144763,0.2028018720426534,0.44921692379616646,0.3686829976188286 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/gpt-neox-20b/results_2024-02-07T16-49-06.259210.json,EleutherAI/gpt-neox-20b,False,20554567680,0.3333333333333333,0.1133653560992191,0.4718588278919615,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/gpt-neox-20b/results_2024-04-04T02-29-53.418614.json,EleutherAI/gpt-neox-20b,False,20554567680,0.3849409111254498,0.24127351840284703,0.4362532523850824,0.3761140819964349 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/polyglot-ko-12.8b/results_2024-02-14T22-19-56.698691.json,EleutherAI/polyglot-ko-12.8b,False,12893603840,0.3333333333333333,0.019117347817184663,0.33515801835418463,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/polyglot-ko-12.8b/results_2025-02-05T22-59-49.919869.json,EleutherAI/polyglot-ko-12.8b,False,12893603840,0.3333333333333333,0.027373568379652526,0.374434139142957,0.33427893671851205 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-12b/results_2024-05-10T12-08-48.948197.json,EleutherAI/pythia-12b,False,11846072320,0.4179270639535547,0.05030336220324859,0.4396551724137931,0.4676874308442238 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-12b-deduped/results_2024-02-14T18-51-32.259622.json,EleutherAI/pythia-12b-deduped,False,11846072320,0.33424046164366045,0.062042189457896954,0.43771626297577854,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-14m/results_2024-02-14T15-03-57.922027.json,EleutherAI/pythia-14m,False,14067712,0.0,,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-14m/results_2024-04-03T19-47-56.339960.json,EleutherAI/pythia-14m,False,14067712,0.33157748821735517,0.0006847937896062521,0.4396551724137931,0.259929067077872 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-160m/results_2024-04-24T22-01-22.569827.json,EleutherAI/pythia-160m,False,162322944,0.5474759773218106,0.05731560767696195,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-160m-deduped/results_2024-02-14T15-27-43.932506.json,EleutherAI/pythia-160m-deduped,False,162322944,0.0,,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-160m-deduped/results_2024-04-03T21-13-20.844629.json,EleutherAI/pythia-160m-deduped,False,162322944,0.3389102997234224,0.04193510248223561,0.4396551724137931,0.45262567913287954 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-1b/results_2024-02-22T09-20-51.293467.json,EleutherAI/pythia-1b,False,1011781632,0.4650037024436128,0.0,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-1b-deduped/results_2024-02-14T16-20-18.432424.json,EleutherAI/pythia-1b-deduped,False,1011781632,0.3333333333333333,0.015287790416728658,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-1b-deduped/results_2024-04-03T22-36-01.685148.json,EleutherAI/pythia-1b-deduped,False,1011781632,0.34088811077510744,0.058535886752163556,0.4396551724137931,0.3503150417064208 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-2.8b/results_2024-04-24T22-06-55.870441.json,EleutherAI/pythia-2.8b,False,2775208960,0.3333333333333333,0.20712471558047654,0.17939674437408695,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-2.8b-deduped/results_2024-02-14T16-49-11.216028.json,EleutherAI/pythia-2.8b-deduped,False,2775208960,0.3333333333333333,0.032750005063908794,0.17721518987341772,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-2.8b-deduped/results_2024-04-03T22-36-58.032425.json,EleutherAI/pythia-2.8b-deduped,False,2775208960,0.3333333333333333,0.25073155863228036,0.17939674437408695,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-410m/results_2024-04-24T22-06-03.860992.json,EleutherAI/pythia-410m,False,405334016,0.3333333333333333,0.029778872408316164,0.4396551724137931,0.34039116046362083 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-410m-deduped/results_2024-02-14T15-42-42.785588.json,EleutherAI/pythia-410m-deduped,False,405334016,0.3333333333333333,0.0393820346422743,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-410m-deduped/results_2024-04-03T21-42-41.391444.json,EleutherAI/pythia-410m-deduped,False,405334016,0.3333333333333333,0.040138096618478045,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-410m-deduped/results_2024-04-03T21-43-53.606908.json,EleutherAI/pythia-410m-deduped,False,405334016,0.3333333333333333,0.040138096618478045,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-6.9b/results_2024-02-22T09-49-44.237199.json,EleutherAI/pythia-6.9b,False,6857302016,0.5918513695309833,0.0025941556675326705,0.3121791039110175,0.32770726983578174 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-6.9b-deduped/results_2024-02-14T17-44-29.767283.json,EleutherAI/pythia-6.9b-deduped,False,6857302016,0.33514630774633175,0.015459703136524651,0.5032594590990455,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-6.9b-deduped/results_2024-04-03T23-20-39.755265.json,EleutherAI/pythia-6.9b-deduped,False,6857302016,0.3333333333333333,0.07157697545169536,0.4396551724137931,0.3349186726374015 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-70m/results_2024-04-24T21-25-37.361813.json,EleutherAI/pythia-70m,False,70426624,0.4502521949740358,0.006173005990956128,0.4396551724137931,0.419144092439547 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-70m-deduped/results_2024-02-14T15-15-51.530711.json,EleutherAI/pythia-70m-deduped,False,70426624,0.0,,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EleutherAI/pythia-70m-deduped/results_2024-04-03T21-10-06.848681.json,EleutherAI/pythia-70m-deduped,False,70426624,0.35073395945309394,0.02026922309956098,0.4138559708295351,0.421145790657915 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EpistemeAI/Fireball-3.1-8B-ORPO/results_2024-09-01T02-54-02.837847.json,EpistemeAI/Fireball-3.1-8B-ORPO,False,8072204288,0.9053984266194903,0.7780445334408981,0.4396551724137931,0.7484501386859586 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EpistemeAI/Fireball-Alpaca-Llama-3.1-8B-Philos-DPO-200/results_2024-09-20T06-15-49.060308.json,EpistemeAI/Fireball-Alpaca-Llama-3.1-8B-Philos-DPO-200,False,8030261248,0.9251856605849454,0.7942568112967181,0.6921595129894318,0.8651061622703166 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EpistemeAI/Fireball-Alpaca-Llama3.1.07-8B-Philos-Math-KTO-beta/results_2024-09-19T02-39-58.689409.json,EpistemeAI/Fireball-Alpaca-Llama3.1.07-8B-Philos-Math-KTO-beta,False,8030261248,0.9276737627600735,0.7920894421060412,0.7165122978572003,0.852197087716136 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EpistemeAI/Fireball-Alpaca-Llama3.1.08-8B-Philos-C-R1-KTO-beta/results_2024-09-20T05-16-46.986428.json,EpistemeAI/Fireball-Alpaca-Llama3.1.08-8B-Philos-C-R1-KTO-beta,False,8030261248,0.9284910083786038,0.7940617995838942,0.7216722083162103,0.8542986033686493 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EpistemeAI/Fireball-Alpaca-Llama3.1.08-8B-Philos-C-R2/results_2024-09-19T01-40-41.608689.json,EpistemeAI/Fireball-Alpaca-Llama3.1.08-8B-Philos-C-R2,False,8030261248,0.8825865492532159,0.7591089850852772,0.3915458645262352,0.8778431197458891 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EpistemeAI/Fireball-Meta-Llama-3.1-8B-Instruct-Agent-0.003-128K-code-ds-auto/results_2024-11-05T01-39-06.988745.json,EpistemeAI/Fireball-Meta-Llama-3.1-8B-Instruct-Agent-0.003-128K-code-ds-auto,False,8030261248,0.8658925885446538,0.7528652824057875,0.5822048202229549,0.8496022132006287 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EpistemeAI/Polypsyche-Llama-3.1-8B-Instruct-Agent-0.003-128K-code-ds-auto-Logic/results_2024-12-21T17-15-42.636701.json,EpistemeAI/Polypsyche-Llama-3.1-8B-Instruct-Agent-0.003-128K-code-ds-auto-Logic,False,8030261248,0.8912897577396475,0.7538649251610948,0.6552756219741444,0.8399736013335878 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EpistemeAI2/Fireball-Alpaca-Llama-3.1-8B-Instruct-KTO-beta/results_2024-09-14T01-41-22.714563.json,EpistemeAI2/Fireball-Alpaca-Llama-3.1-8B-Instruct-KTO-beta,False,8030261248,0.9289006399663491,0.790665703097916,0.7262640413225272,0.8542568542568543 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EpistemeAI2/Fireball-Alpaca-Llama3.1.06-8B-Philos/results_2025-02-04T17-42-27.526477.json,EpistemeAI2/Fireball-Alpaca-Llama3.1.06-8B-Philos,False,8030261248,0.9227291181885153,0.7965276873341587,0.6093005998392184,0.8686810440933379 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/EpistemeAI2/Fireball-Llama-3.1-8B-Philos-Relection/results_2024-09-20T07-15-00.048912.json,EpistemeAI2/Fireball-Llama-3.1-8B-Philos-Relection,False,8030261248,0.926867357294864,0.759448794502228,0.7867783266440259,0.8599263715542785 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ericzzz/falcon-rw-1b-instruct-openorca/results_2024-04-23T00-40-09.642203.json,ericzzz/falcon-rw-1b-instruct-openorca,False,1311625216,0.3428635121469901,0.05037353224367054,0.4396551724137931,0.35268354828334864 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Eurdem/Defne-llama3.1-8B/results_2024-08-13T21-37-00.956954.json,Eurdem/Defne-llama3.1-8B,False,8030261248,0.9313503711043392,0.7443820349848373,0.7271022126258165,0.8745591008396347 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/galactica-1.3b/results_2024-05-10T17-37-38.266579.json,facebook/galactica-1.3b,False,1315201024,0.36988739222484734,0.00014177526663868966,0.4396551724137931,0.3330157217722725 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/galactica-125m/results_2024-05-10T16-55-49.372246.json,facebook/galactica-125m,False,125030400,0.5910688252686159,0.0648907903481681,0.4396551724137931,0.3649606106353511 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/galactica-30b/results_2024-05-21T20-48-11.074507.json,facebook/galactica-30b,False,29972590592,0.51833455248139,0.14743934926911426,0.5034538805363673,0.40873283299191565 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/galactica-6.7b/results_2024-05-22T06-41-07.644691.json,facebook/galactica-6.7b,False,6657359872,0.47661462677231026,0.07886060654880883,0.4396551724137931,0.44181680655420985 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-1.3b/results_2024-02-07T17-55-47.448141.json,facebook/opt-1.3b,False,1315758080,0.3333333333333333,0.014294092039832406,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-1.3b/results_2024-04-04T05-03-48.444586.json,facebook/opt-1.3b,False,1315758080,0.3674033988994619,0.045560010081070086,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-125m/results_2024-02-07T17-00-47.502507.json,facebook/opt-125m,False,125239296,0.3333333333333333,0.045915967954967385,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-125m/results_2024-04-04T03-10-11.741276.json,facebook/opt-125m,False,125239296,0.43316031183469134,0.023023204939196636,0.4396551724137931,0.3728445198230819 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-13b/results_2024-02-07T19-29-22.779324.json,facebook/opt-13b,False,12853473280,0.3333333333333333,0.013938417309944717,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-13b/results_2024-04-04T07-01-31.332443.json,facebook/opt-13b,False,12853473280,0.4665309623379448,0.020511169211933078,0.4396551724137931,0.5714040447709018 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-2.7b/results_2024-02-07T18-37-10.420552.json,facebook/opt-2.7b,False,2651596800,0.3333333333333333,0.016315511418924157,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-2.7b/results_2024-04-04T06-39-26.624482.json,facebook/opt-2.7b,False,2651596800,0.3333333333333333,0.019515420426272946,0.4396551724137931,0.4603905744413567 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-30b/results_2024-02-07T20-24-35.022523.json,facebook/opt-30b,False,29974540288,0.3600025224060601,0.04892185781466516,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-30b/results_2024-02-20T17-45-32.313009.json,facebook/opt-30b,False,29974540288,0.6606852984343372,0.016937221060787003,0.4396551724137931,0.6952554678480041 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-350m/results_2024-02-07T17-20-17.729922.json,facebook/opt-350m,False,331196416,0.3333333333333333,0.018818660102831564,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-350m/results_2024-04-04T03-52-34.527925.json,facebook/opt-350m,False,331196416,0.3333333333333333,0.13015841245703968,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-6.7b/results_2024-02-07T19-21-03.591228.json,facebook/opt-6.7b,False,6658473984,0.3333333333333333,0.14409507648515532,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-6.7b/results_2024-02-22T12-20-35.811780.json,facebook/opt-6.7b,False,6658473984,0.3913233591528234,0.0754529305698404,0.4396551724137931,0.36628945801072066 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-66b/results_2024-02-12T22-14-09.097826.json,facebook/opt-66b,False,65719701504,0.48633439489737007,0.03942223193215699,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/opt-66b/results_2024-02-20T21-20-17.294248.json,facebook/opt-66b,False,65719701504,0.5553268569422898,0.10541118742184293,0.446059558419109,0.6215758016105734 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/xglm-1.7B/results_2024-05-21T22-30-52.997014.json,facebook/xglm-1.7B,False,1732907008,0.33605087493777575,0.07477282113396852,0.4396551724137931,0.37168470829147143 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/xglm-1.7B/results_2024-05-21T22-30-53.852635.json,facebook/xglm-1.7B,False,1732907008,0.33605087493777575,0.07477282113396852,0.4396551724137931,0.37168470829147143 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/xglm-2.9B/results_2024-05-21T22-30-53.060182.json,facebook/xglm-2.9B,False,2941505536,0.3333333333333333,0.05617638414772318,0.4396551724137931,0.32752350236739175 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/xglm-2.9B/results_2024-05-21T22-30-53.897206.json,facebook/xglm-2.9B,False,2941505536,0.3333333333333333,0.05617638414772318,0.4396551724137931,0.32752350236739175 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/xglm-4.5B/results_2024-05-22T07-29-35.683620.json,facebook/xglm-4.5B,False,4552511488,0.3333333333333333,0.04298399826502508,0.43722943722943725,0.3394871794871795 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/xglm-564M/results_2024-05-21T22-03-00.248630.json,facebook/xglm-564M,False,564463616,0.49898320559874154,0.23347584115631292,0.4396551724137931,0.3398976841410418 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/facebook/xglm-7.5B/results_2024-05-22T07-58-52.741101.json,facebook/xglm-7.5B,False,7492771840,0.5107106652146206,0.05562042288433293,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/failspy/Codestral-22B-v0.1-abliterated-v3/results_2024-07-03T04-04-20.379881.json,failspy/Codestral-22B-v0.1-abliterated-v3,False,22247282688,0.8825766952657541,0.796870925932544,0.566146041916967,0.8209886168719789 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/failspy/Meta-Llama-3-8B-Instruct-abliterated-v3/results_2024-05-27T12-30-50.126635.json,failspy/Meta-Llama-3-8B-Instruct-abliterated-v3,False,8030261248,0.9073768010736127,0.7546147599967769,0.755329706053451,0.869595212895539 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/failspy/Phi-3-medium-4k-instruct-abliterated-v3/results_2024-05-25T08-02-58.396353.json,failspy/Phi-3-medium-4k-instruct-abliterated-v3,False,13960238080,0.9256174008161892,0.8022099861362467,0.7299937983407636,0.7935385931129845 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/fblgit/TheBeagle-v2beta-32B-MGS/results_2024-12-21T14-21-43.929072.json,fblgit/TheBeagle-v2beta-32B-MGS,False,32763876352,0.9472943438455241,0.7889619626227551,0.8226761081260601,0.9191698715148393 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/fblgit/UNA-SimpleSmaug-34b-v1beta/results_2024-06-16T10-06-17.462875.json,fblgit/UNA-SimpleSmaug-34b-v1beta,False,34388917248,0.917886114144189,0.8082048086835643,0.8097848807348216,0.6897101227639736 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/fernandosola/bluearara-7B/results_2024-02-21T13-54-52.697317.json,fernandosola/bluearara-7B,False,7241732096,0.7161848185259556,0.1361071782410529,0.4396551724137931,0.6226564096049223 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/fernandosola/bluearara-7B/results_2024-02-22T13-31-58.902845.json,fernandosola/bluearara-7B,False,7241732096,0.7104594402253265,0.1302382514093839,0.4396551724137931,0.5745222209657356 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/fernandosola/bluearara-7B-instruct/results_2024-03-09T01-52-25.234067.json,fernandosola/bluearara-7B-instruct,False,7241732096,0.5068425347237789,0.17174767635340848,0.4439232761621561,0.8426873093605136 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/fernandosola/bluearara-7B-instruct-v02/results_2025-01-31T00-26-58.153487.json,fernandosola/bluearara-7B-instruct-v02,False,7241732096,0.5604414655666146,0.21301527952103266,0.45333525567089583,0.8413894043222918 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/FluxiIA/Tucaninho/results_2025-04-03T22-25-00.182981.json,FluxiIA/Tucaninho,False,2444618240,0.3333333333333333,0.11238071919631473,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/FluxiIA/Tucaninho_dpo/results_2025-04-03T23-17-58.127114.json,FluxiIA/Tucaninho_dpo,False,2444618240,0.3333333333333333,0.11564508409018462,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/freewheelin/free-llama3-dpo-v0.2/results_2024-08-12T04-15-46.815373.json,freewheelin/free-llama3-dpo-v0.2,False,8030261248,0.895435486975249,0.7765266903702241,0.4471267110923455,0.7581040751599846 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/freewheelin/free-llama3-dpo-v0.2/results_2024-08-12T04-18-54.904777.json,freewheelin/free-llama3-dpo-v0.2,False,8030261248,0.895435486975249,0.7765266903702241,0.4471267110923455,0.7581040751599846 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/freewheelin/free-solar-evo-v0.1/results_2024-08-11T12-35-46.464008.json,freewheelin/free-solar-evo-v0.1,False,10731524096,0.6023639703251354,0.655402578466106,0.3555050872093023,0.34136229074155156 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/freewheelin/free-solar-evo-v0.11/results_2024-08-09T06-28-11.333017.json,freewheelin/free-solar-evo-v0.11,False,10731524096,0.649323259668555,0.6657294593588593,0.39924242424242423,0.315487166961796 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/freewheelin/free-solar-evo-v0.13/results_2024-08-12T03-39-46.797616.json,freewheelin/free-solar-evo-v0.13,False,10731524096,0.6832664866290573,0.6703707809587911,0.4538823122514109,0.40663023261530395 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/FuseAI/FuseChat-7B-VaRM/results_2024-03-08T15-26-39.517660.json,FuseAI/FuseChat-7B-VaRM,False,7241748480,0.9272868051476477,0.7836651113903375,0.787259111855886,0.8223021238433512 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/FuseAI/OpenChat-3.5-7B-Solar/results_2024-03-08T13-22-15.392524.json,FuseAI/OpenChat-3.5-7B-Solar,False,7241748480,0.927694039333938,0.7822958272680564,0.7777514761773254,0.8223021238433512 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/gaverfraxz/Meta-Llama-3.1-8B-Instruct-HalfAbliterated-TIES/results_2025-02-04T21-05-37.086653.json,gaverfraxz/Meta-Llama-3.1-8B-Instruct-HalfAbliterated-TIES,False,8030261248,0.9068582926802695,0.7439850054712727,0.7778893287086737,0.8607142146501094 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/GeneZC/MiniChat-1.5-3B/results_2024-07-25T12-46-20.562745.json,GeneZC/MiniChat-1.5-3B,False,3020442624,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/GeneZC/MiniChat-2-3B/results_2024-07-29T03-48-21.141755.json,GeneZC/MiniChat-2-3B,False,3020442624,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/GeneZC/MiniChat-3B/results_2024-07-25T10-05-31.284967.json,GeneZC/MiniChat-3B,False,3020442624,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/GeneZC/MiniLoong-3B/results_2024-07-29T03-08-19.036276.json,GeneZC/MiniLoong-3B,False,3020442624,0.33424046164366045,0.21769252072678535,0.5193010971149938,0.7281872301274692 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/GeneZC/MiniMA-2-1B/results_2024-07-26T02-13-22.435122.json,GeneZC/MiniMA-2-1B,False,1373490432,0.6731362234717939,0.1928558925744654,0.4386873920552677,0.5942961348427174 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/GeneZC/MiniMA-2-3B/results_2024-07-26T02-44-22.514782.json,GeneZC/MiniMA-2-3B,False,3020442624,0.3649999400052577,0.014848574812705627,0.334319526627219,0.5384799808199472 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/GeneZC/MiniMA-3B/results_2024-07-26T01-32-17.409993.json,GeneZC/MiniMA-3B,False,3020442624,0.5688044306828837,0.03789027230859506,0.45243742933782616,0.48903191180968997 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ghost-x/ghost-8b-beta/results_2024-07-24T08-26-58.470678.json,ghost-x/ghost-8b-beta,False,8030261248,0.9157033248081841,0.6786078768211791,0.608080080260396,0.8476219675775112 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ghost-x/ghost-8b-beta-1608/results_2024-08-25T01-39-27.386034.json,ghost-x/ghost-8b-beta-1608,False,8030261248,0.9065210692034569,0.6934426233290423,0.363744357329532,0.8457130262287855 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Goekdeniz-Guelmez/Josiefied-Qwen2.5-14B-Instruct-abliterated-v4/results_2024-11-02T02-40-08.132067.json,Goekdeniz-Guelmez/Josiefied-Qwen2.5-14B-Instruct-abliterated-v4,False,14770033664,0.9436108523871045,0.8625851172324958,0.7810834599851131,0.9112762925756247 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-1.1-2b-it/results_2024-04-11T20-03-36.311842.json,google/gemma-1.1-2b-it,False,2506172416,0.5921706256532004,0.012885843184192301,0.4510135135135135,0.5459354845570212 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-1.1-7b-it/results_2024-04-13T10-03-56.394429.json,google/gemma-1.1-7b-it,False,8537680896,0.868278010240231,0.3852430157760904,0.6359748948203324,0.6711070120333069 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-2-27b/results_2024-07-05T20-07-07.757412.json,google/gemma-2-27b,False,27227128320,0.8990786820541422,0.7868546967741642,0.580029368575624,0.7416170395407479 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-2-27b-it/results_2024-07-05T22-22-00.449825.json,google/gemma-2-27b-it,False,27227128320,0.9238304264765416,0.7633684342010387,0.7871649805957103,0.8829863003724235 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-2-2b/results_2024-08-11T17-38-07.802583.json,google/gemma-2-2b,False,2614341888,0.7444676409185804,0.4655280185096118,0.45449901481427424,0.6618591929795528 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-2-2b/results_2024-08-11T17-38-07.811610.json,google/gemma-2-2b,False,2614341888,0.7444676409185804,0.4655280185096118,0.45449901481427424,0.6618591929795528 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-2-2b-it/results_2024-08-11T16-36-29.756962.json,google/gemma-2-2b-it,False,2614341888,0.8878200067968414,0.7220392537996184,0.7274889408703897,0.8818036456667733 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-2-2b-it/results_2024-08-13T17-48-07.559693.json,google/gemma-2-2b-it,False,2614341888,0.8853170946194202,0.7207393337752955,0.7325826055270179,0.8775597254351002 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-2-9b/results_2024-07-04T19-32-15.424185.json,google/gemma-2-9b,False,9241705984,0.8803683624255816,0.7487487401357935,0.48227339504752464,0.7531373915137753 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-2-9b-it/results_2024-07-05T05-56-40.657388.json,google/gemma-2-9b-it,False,9241705984,0.9390883593858939,0.7866160922637744,0.7822847968953928,0.887461593324701 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-2b/results_2024-02-22T03-02-32.420657.json,google/gemma-2b,False,2506172416,0.6494380403222241,0.3568300099073752,0.4486979644232532,0.7809708451948563 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-2b/results_2024-04-04T14-48-26.410309.json,google/gemma-2b,False,2506172416,0.6353400730210679,0.3635210295220489,0.44754198680900253,0.7782210622319081 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-2b-it/results_2024-02-22T03-31-32.102531.json,google/gemma-2b-it,False,2506172416,0.4559526524168822,0.04640114163897497,0.5461440300483953,0.4030721280250801 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-2b-it/results_2024-02-23T23-17-06.001747.json,google/gemma-2b-it,False,2506172416,0.5748320775250757,0.03912102519009544,0.5581954887218045,0.44155722702435507 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-2b-it/results_2024-04-04T15-40-56.547837.json,google/gemma-2b-it,False,2506172416,0.5717359986097028,0.05512612934965275,0.5519874376678174,0.4455445544554455 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-3-12b-it/results_2025-04-06T08-25-09.008958.json,google/gemma-3-12b-it,False,12187325040,0.9386425568007508,0.832205538885632,0.7661029147175242,0.8619656524703967 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-3-12b-pt/results_2025-04-07T02-39-19.962477.json,google/gemma-3-12b-pt,False,12187325040,0.9140605775300374,0.7592033553714468,0.4396551724137931,0.776823418127766 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-3-1b-it/results_2025-04-06T07-15-58.657956.json,google/gemma-3-1b-it,False,999885952,0.8167410548701626,0.4417871177368796,0.5763183990907976,0.8264086438494215 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-3-1b-pt/results_2025-04-06T07-35-36.577757.json,google/gemma-3-1b-pt,False,999885952,0.45828659574880215,0.07222305923135837,0.4396551724137931,0.6979988739126185 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-3-4b-it/results_2025-04-06T07-53-32.688856.json,google/gemma-3-4b-it,False,4300079472,0.904309570646596,0.7193038819802234,0.6942389006342495,0.8390820356895634 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-3n-E2B/results_2025-07-15T17-03-06.578799.json,google/gemma-3n-E2B,False,5439438208,0.8686747734963725,0.0,0.17721518987341772,0.5689872592077228 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-3n-E2B-it/results_2025-07-15T16-35-32.997841.json,google/gemma-3n-E2B-it,False,5439438208,0.9152606173812926,0.7120974389960102,0.66221834206369,0.8007176057118515 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-3n-E4B/results_2025-07-15T18-01-29.832699.json,google/gemma-3n-E4B,False,7849978128,0.8927073749293221,0.0,0.4214891683078321,0.7895308715702862 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-3n-E4B-it/results_2025-07-15T17-28-52.132900.json,google/gemma-3n-E4B-it,False,7849978128,0.9218387538145081,0.7529776546800084,0.737833716407355,0.8255582137161085 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-7b/results_2024-02-21T20-58-48.539461.json,google/gemma-7b,False,8537680896,0.8137031006290386,0.6382518953799198,0.6854007277477141,0.8395989974937343 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-7b/results_2024-04-04T16-10-55.518267.json,google/gemma-7b,False,8537680896,0.8134100000770779,0.6427577701951299,0.6923244279409992,0.8568972460897173 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-7b-it/results_2024-02-22T01-58-50.802183.json,google/gemma-7b-it,False,8537680896,0.3770055461311309,0.03879981639793527,0.43364270050125314,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-7b-it/results_2024-02-24T01-02-07.220409.json,google/gemma-7b-it,False,8537680896,0.8049543946616404,0.5173574778975697,0.5420636225946374,0.7699831007992424 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-7b-it/results_2024-02-24T02-28-13.456002.json,google/gemma-7b-it,False,8537680896,0.8148198753180605,0.6045013636313136,0.5736061452793719,0.728133553621187 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-7b-it/results_2024-04-04T18-32-09.431896.json,google/gemma-7b-it,False,8537680896,0.8062892945644999,0.5296649587496869,0.5617861675126903,0.7656970716215885 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-7b-it/results_2024-04-04T20-42-51.153597.json,google/gemma-7b-it,False,8537680896,0.8062892945644999,0.5296649587496869,0.5617861675126903,0.7656970716215885 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/gemma-7b-it/results_2024-04-04T20-51-23.305083.json,google/gemma-7b-it,False,8537680896,0.8144223609599633,0.6083956774707528,0.5736061452793719,0.7273155692737169 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/medgemma-27b-it/results_2025-07-15T21-16-35.431035.json,google/medgemma-27b-it,False,27432406640,0.9460740760519473,0.7871083311853868,0.8159326328340413,0.875901875901876 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/medgemma-27b-text-it/results_2025-07-15T18-33-23.044428.json,google/medgemma-27b-text-it,False,27009002240,0.9481103079890625,0.8190445432478614,0.8520603980183263,0.844150223370808 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/medgemma-4b-it/results_2025-07-15T16-11-03.530001.json,google/medgemma-4b-it,False,4300079472,0.9071887095764003,0.7161228149892342,0.592948717948718,0.8481514208945495 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/medgemma-4b-pt/results_2025-07-15T15-46-49.329632.json,google/medgemma-4b-pt,False,4300079472,0.7976256093464154,0.0,0.5248252518576378,0.7096104392549347 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/mt5-base/results_2024-04-18T06-10-32.403150.json,google/mt5-base,False,582401280,0.0008163265306122449,0.0435769262346769,0.34021739130434786,0.007032348804500704 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/mt5-base/results_2024-04-19T18-04-31.312147.json,google/mt5-base,False,582401280,0.0,0.030671454944035954,0.3625128733264676,0.004249291784702549 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/mt5-small/results_2024-04-18T05-31-05.645206.json,google/mt5-small,False,300176768,0.0008163265306122449,0.07222006970775031,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/mt5-small/results_2024-04-19T16-59-17.995892.json,google/mt5-small,False,300176768,0.0,0.07821073867599507,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/recurrentgemma-2b/results_2024-05-18T17-19-05.513357.json,google/recurrentgemma-2b,False,2682862080,0.7512533304333642,0.13792975851339373,0.4396551724137931,0.6965738263691794 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/recurrentgemma-2b-it/results_2024-05-18T03-42-04.057438.json,google/recurrentgemma-2b-it,False,2682862080,0.7942082397406754,0.3566309725066242,0.281968031968032,0.7547466132100767 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/umt5-base/results_2024-02-15T05-17-17.944550.json,google/umt5-base,False,592043520,0.008049597487606372,0.11702145458274152,0.005177993527508091,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/google/umt5-small/results_2024-02-15T02-03-26.881523.json,google/umt5-small,False,306601984,0.19257980378257433,0.043458369255268595,0.014741228752745068,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/gpt2/results_2024-02-14T01-24-06.058943.json,gpt2,False,124439808,0.3333333333333333,0.039726178664430394,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/gpt2/results_2024-04-02T10-05-30.859827.json,gpt2,False,124439808,0.3333333333333333,0.03315295501483895,0.4518214740462686,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/gradientai/Llama-3-8B-Instruct-262k/results_2024-06-15T16-29-28.786440.json,gradientai/Llama-3-8B-Instruct-262k,False,8030261248,0.8772215291835841,0.733909567114553,0.40769090577729183,0.8569300040263985 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/grimjim/Llama-3-Instruct-8B-SimPO-SPPO-Iter3-merge/results_2024-09-20T08-14-17.032678.json,grimjim/Llama-3-Instruct-8B-SimPO-SPPO-Iter3-merge,False,8030261248,0.910441163867953,0.7115193436394658,0.6897004415801409,0.8540491600637007 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/grimjim/Llama-3-Instruct-8B-SPPO-Iter3-SimPO-merge/results_2024-08-11T16-25-38.137060.json,grimjim/Llama-3-Instruct-8B-SPPO-Iter3-SimPO-merge,False,8030261248,0.9129288088010328,0.7129374319103543,0.7097079326156245,0.8569250386831352 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/grimjim/Llama-3.1-8B-Instruct-abliterated_via_adapter/results_2024-10-06T01-44-53.133934.json,grimjim/Llama-3.1-8B-Instruct-abliterated_via_adapter,False,8030261248,0.9169043186698912,0.7642040160077297,0.7277384958072006,0.8761935286326745 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/grimjim/Llama-3.1-8B-Instruct-abliterated_via_adapter/results_2024-10-06T01-44-53.643982.json,grimjim/Llama-3.1-8B-Instruct-abliterated_via_adapter,False,8030261248,0.9169043186698912,0.7642040160077297,0.7277384958072006,0.8761935286326745 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/GritLM/GritLM-7B/results_2024-06-12T20-31-09.833902.json,GritLM/GritLM-7B,False,7241732096,0.9128900575777901,0.7210345169515866,0.8085014118262378,0.8066603535353536 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/GritLM/GritLM-7B-KTO/results_2024-06-15T00-03-38.180499.json,GritLM/GritLM-7B-KTO,False,7241732096,0.9203424859265217,0.7712035054624826,0.7270998493492381,0.7778660429262838 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/GuiCas/Phituguese_FP16/results_2024-06-11T22-22-16.921558.json,GuiCas/Phituguese_FP16,False,3821079552,0.7475183753883459,0.7193372156199731,0.4396551724137931,0.5733866698041061 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/h2oai/h2o-danube-1.8b-base/results_2024-04-12T03-45-40.969245.json,h2oai/h2o-danube-1.8b-base,False,1831201280,0.5377611637321853,0.028670997778142325,0.4396551724137931,0.6997952571886329 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/h2oai/h2o-danube-1.8b-chat/results_2024-04-12T02-52-47.488340.json,h2oai/h2o-danube-1.8b-chat,False,1831201280,0.5133751405236282,0.23691559149369143,0.4471267110923455,0.636027706300527 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/h2oai/h2o-danube-1.8b-sft/results_2024-05-21T23-57-17.751493.json,h2oai/h2o-danube-1.8b-sft,False,1831201280,0.6036670511808633,0.08561453836574923,0.4396551724137931,0.5955978159794826 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/h2oai/h2o-danube2-1.8b-base/results_2024-04-12T00-04-18.015443.json,h2oai/h2o-danube2-1.8b-base,False,1831201280,0.6292397280327922,0.26950270669436893,0.4396551724137931,0.8164863143512329 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/h2oai/h2o-danube2-1.8b-chat/results_2024-04-12T02-23-05.169659.json,h2oai/h2o-danube2-1.8b-chat,False,1831201280,0.6528434251160538,0.19401751168680534,0.571463395716296,0.8305959621797061 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/h2oai/h2o-danube2-1.8b-sft/results_2024-05-22T00-26-40.598093.json,h2oai/h2o-danube2-1.8b-sft,False,1831201280,0.7062162848305584,0.1922050515236594,0.5454568322340784,0.769365386386663 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/h2oai/h2o-danube2-1.8b-sft/results_2024-05-22T00-27-53.218756.json,h2oai/h2o-danube2-1.8b-sft,False,1831201280,0.7062162848305584,0.1922050515236594,0.5454568322340784,0.769365386386663 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/h2oai/h2o-danube3-4b-base/results_2024-07-24T10-24-17.863830.json,h2oai/h2o-danube3-4b-base,False,3961839360,0.7647409015101712,0.5132955743220948,0.5432549748578435,0.640884477350184 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/h2oai/h2o-danube3-4b-chat/results_2024-12-05T18-54-22.927528.json,h2oai/h2o-danube3-4b-chat,False,3961839360,0.7691750586979651,0.5995060623128778,0.5119463869463869,0.7753225829982973 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/h2oai/h2o-danube3-500m-base/results_2024-07-24T09-56-11.157649.json,h2oai/h2o-danube3-500m-base,False,513590784,0.47838248555653184,0.10337592586940014,0.4396551724137931,0.48458793790226434 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/h2oai/h2o-danube3-500m-chat/results_2024-07-24T13-20-20.190488.json,h2oai/h2o-danube3-500m-chat,False,513590784,0.3376711508522614,0.015297942146281086,0.4396551724137931,0.3607305936073059 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/h2oai/h2o-danube3.1-4b-chat/results_2024-12-05T17-51-52.208655.json,h2oai/h2o-danube3.1-4b-chat,False,3961839360,0.7579152162104068,0.5451955916078189,0.5758841184914525,0.8452338790408594 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/haoranxu/ALMA-13B-Pretrain/results_2024-05-21T01-04-37.356835.json,haoranxu/ALMA-13B-Pretrain,False,13015864320,0.7875239701530645,0.26615781879196837,0.43917169974115616,0.6435130415187877 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/haoranxu/ALMA-13B-R/results_2024-05-19T06-12-48.513901.json,haoranxu/ALMA-13B-R,False,13027988480,0.7857737944388717,0.16284027859513017,0.4396551724137931,0.5900914612050847 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/haoranxu/ALMA-7B-R/results_2024-08-07T08-00-47.749569.json,haoranxu/ALMA-7B-R,False,6746148864,0.5999564976704201,0.11974200021896443,0.4442339269925477,0.7559659680431625 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/heloisy/zephyr-7b-b-cosmosqa-ptbr-text-v2/results_2024-05-28T22-07-52.653697.json,heloisy/zephyr-7b-b-cosmosqa-ptbr-text-v2,False,7306743808,0.8868409480985546,0.6755767261277698,0.6476497073888465,0.7530267389366248 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/heloisy/zephyr-7b-b-cosmosqa-ptbr-text-v3/results_2024-05-28T01-03-37.730666.json,heloisy/zephyr-7b-b-cosmosqa-ptbr-text-v3,False,7274237952,0.8905017534700548,0.6749346523503066,0.645511263652361,0.7714165715110708 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/heloisy/zephyr-7b-cosmosqa-ptbr_text/results_2024-05-27T23-27-46.657807.json,heloisy/zephyr-7b-cosmosqa-ptbr_text,False,7274237952,0.8950123980409443,0.6847128283716246,0.6308160668847691,0.7842785385447502 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HeyLucasLeao/gpt-neo-small-portuguese/results_2024-02-08T01-23-52.798469.json,HeyLucasLeao/gpt-neo-small-portuguese,False,125199360,0.3333333333333333,0.04753851053349196,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HeyLucasLeao/gpt-neo-small-portuguese/results_2024-03-08T04-18-26.971751.json,HeyLucasLeao/gpt-neo-small-portuguese,False,125199360,0.3528931097729911,0.040770337667175804,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/hkust-nlp/deita-7b-v1.0/results_2024-05-20T23-09-35.076966.json,hkust-nlp/deita-7b-v1.0,False,7241732096,0.8520896848392872,0.7580664670996959,0.7632014424026872,0.8462835603587262 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/hkust-nlp/deita-7b-v1.0-sft/results_2024-04-26T02-53-48.322939.json,hkust-nlp/deita-7b-v1.0-sft,False,7241732096,0.8569157392686804,0.71843542014848,0.6364093293231621,0.71177122003125 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/hotmailuser/QwenSlerp2-14B/results_2025-01-31T14-02-00.942614.json,hotmailuser/QwenSlerp2-14B,False,14765947904,0.9432030222551308,0.8550225495859294,0.8022489790664749,0.8960453713305424 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HuggingFaceH4/zephyr-7b-alpha/results_2024-04-15T10-11-49.222023.json,HuggingFaceH4/zephyr-7b-alpha,False,7241732096,0.9011395676691729,0.7233470427220756,0.6962929525710168,0.8526041634724087 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HuggingFaceH4/zephyr-7b-beta/results_2024-02-21T23-57-52.146406.json,HuggingFaceH4/zephyr-7b-beta,False,7241732096,0.8836486323653452,0.6678266192299295,0.7017672651113582,0.8176778106453834 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HuggingFaceH4/zephyr-7b-gemma-v0.1/results_2024-03-08T19-16-54.419572.json,HuggingFaceH4/zephyr-7b-gemma-v0.1,False,8537680896,0.859763198757764,0.7243826042545414,0.7508333333333334,0.8740224469094597 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HuggingFaceH4/zephyr-7b-gemma-v0.1/results_2024-04-04T23-04-13.841492.json,HuggingFaceH4/zephyr-7b-gemma-v0.1,False,8537680896,0.8604729280813948,0.7259016112950178,0.7486076732673268,0.8755151098901099 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HuggingFaceTB/cosmo-1b/results_2024-02-26T18-03-06.542808.json,HuggingFaceTB/cosmo-1b,False,1741785088,0.5526600270022243,0.07330211383402985,0.4396551724137931,0.3411338879766024 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HuggingFaceTB/SmolLM-1.7B/results_2024-07-24T12-31-42.936886.json,HuggingFaceTB/SmolLM-1.7B,False,1711376384,0.40582136458425117,0.02207394442525775,0.4396551724137931,0.33650009913117634 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HuggingFaceTB/SmolLM-1.7B-Instruct/results_2024-07-29T01-29-23.830440.json,HuggingFaceTB/SmolLM-1.7B-Instruct,False,1711376384,0.5953496404707331,0.011089425612064035,0.26732673267326734,0.3739763527924611 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HuggingFaceTB/SmolLM-135M/results_2024-07-24T11-11-26.158993.json,HuggingFaceTB/SmolLM-135M,False,134515008,0.44365759807918503,0.0402831387140938,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HuggingFaceTB/SmolLM-135M-Instruct/results_2024-07-24T14-22-33.098781.json,HuggingFaceTB/SmolLM-135M-Instruct,False,134515008,0.3945071581566081,0.03362436337787243,0.10587254318046457,0.09975522704208357 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HuggingFaceTB/SmolLM-360M/results_2024-07-24T11-50-44.598786.json,HuggingFaceTB/SmolLM-360M,False,361821120,0.33514630774633175,0.10461683763578711,0.4396551724137931,0.42822089483267245 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HuggingFaceTB/SmolLM-360M-Instruct/results_2024-07-24T17-08-31.061263.json,HuggingFaceTB/SmolLM-360M-Instruct,False,361821120,0.3333333333333333,0.036062757273046704,0.13893376413570277,0.3523595573670525 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HuggingFaceTB/SmolLM3-3B/results_2025-07-09T03-37-09.519440.json,HuggingFaceTB/SmolLM3-3B,False,3075098624,0.927284475393028,0.7417965137337437,0.7262640413225272,0.7894431324951516 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/HuggingFaceTB/SmolLM3-3B-Base/results_2025-07-09T03-13-54.804416.json,HuggingFaceTB/SmolLM3-3B-Base,False,3075098624,0.8709108106322094,0.0,0.5519970216983363,0.7549056237203533 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huggyllama/llama-13b/results_2024-02-06T03-05-10.640098.json,huggyllama/llama-13b,False,13015864320,0.5505784757177058,0.3575846951731686,0.5734565832149147,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huggyllama/llama-13b/results_2024-02-16T20-08-01.336839.json,huggyllama/llama-13b,False,13015864320,0.7971928190943076,0.32144695937781287,0.5488046385333089,0.8226847619140406 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huggyllama/llama-30b/results_2024-02-10T05-59-10.907847.json,huggyllama/llama-30b,False,32528943616,0.6994823029869264,0.521939545377829,0.5100755946706865,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huggyllama/llama-30b/results_2024-03-04T06-00-26.904208.json,huggyllama/llama-30b,False,32528943616,0.7293812533558872,0.6049405695262198,0.519508339133766,0.6823108859287468 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huggyllama/llama-65b/results_2024-02-10T12-55-13.380671.json,huggyllama/llama-65b,False,65285660672,0.7283405417709993,0.254385232549328,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huggyllama/llama-65b/results_2024-03-03T23-57-44.587602.json,huggyllama/llama-65b,False,65285660672,0.7707707349316295,0.6875101964208782,0.4396551724137931,0.6774236636123208 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huggyllama/llama-7b/results_2024-02-05T23-45-55.633251.json,huggyllama/llama-7b,False,6738415616,0.3333333333333333,0.07703722378109175,0.5548892542799111,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huggyllama/llama-7b/results_2024-02-06T03-05-10.639848.json,huggyllama/llama-7b,False,6738415616,0.3333333333333333,0.07703722378109175,0.5548892542799111,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huggyllama/llama-7b/results_2024-02-16T19-03-42.667513.json,huggyllama/llama-7b,False,6738415616,0.38621506701188596,0.15427900101339342,0.5242660399142339,0.3807353552771385 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huihui-ai/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2/results_2025-01-31T22-20-22.473642.json,huihui-ai/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2,False,14770033664,0.8968490072231892,0.004219444060059984,0.0,0.9178477121098085 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huihui-ai/Qwen2.5-32B-Instruct-abliterated/results_2024-12-03T07-52-40.624605.json,huihui-ai/Qwen2.5-32B-Instruct-abliterated,False,32763876352,0.9448208891203416,0.7909383241753514,0.7829996268771865,0.9274548520913756 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huihui-ai/Qwen2.5-7B-Instruct-abliterated/results_2024-09-24T03-29-31.645903.json,huihui-ai/Qwen2.5-7B-Instruct-abliterated,False,7615616512,0.9411754889860902,0.7622235710591121,0.8153773957715866,0.7593890574327858 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huihui-ai/Qwen2.5-7B-Instruct-abliterated-v2/results_2024-10-05T07-16-03.986707.json,huihui-ai/Qwen2.5-7B-Instruct-abliterated-v2,False,7615616512,0.9383157479837994,0.7407777788633316,0.8544933327542024,0.771629617783464 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huihui-ai/Qwen2.5-7B-Instruct-abliterated-v2/results_2024-10-05T07-19-15.271156.json,huihui-ai/Qwen2.5-7B-Instruct-abliterated-v2,False,7615616512,0.9383157479837994,0.7407777788633316,0.8544933327542024,0.771629617783464 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huihui-ai/Qwen2.5-Coder-32B-Instruct-abliterated/results_2024-12-21T11-32-43.628868.json,huihui-ai/Qwen2.5-Coder-32B-Instruct-abliterated,False,32763876352,0.9321779534287545,0.7886416189551377,0.799725388957525,0.7908076819305252 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huihui-ai/QwQ-32B-Coder-Fusion-9010/results_2024-12-21T08-42-09.851644.json,huihui-ai/QwQ-32B-Coder-Fusion-9010,False,32763876352,0.8529472776210971,0.8292057259952751,0.6617496617496618,0.9205098418943825 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/huihui-ai/QwQ-32B-Preview-abliterated/results_2024-12-21T05-50-27.527835.json,huihui-ai/QwQ-32B-Preview-abliterated,False,32763876352,0.8994699769154177,0.8329921858686682,0.7217230443974629,0.9248840685241253 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ibivibiv/llama-3-nectar-dpo-8B/results_2024-06-22T07-38-35.422279.json,ibivibiv/llama-3-nectar-dpo-8B,False,8030261248,0.9082379542568665,0.7521023848238035,0.751649303344456,0.8684745291939572 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ibivibiv/multimaster-7b-v6/results_2024-06-14T13-39-45.160271.json,ibivibiv/multimaster-7b-v6,False,35428241408,0.9219507504908876,0.775788909453694,0.7186147186147187,0.867516403157304 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/informatiker/Qwen2-7B-Instruct-abliterated/results_2024-09-19T01-44-47.552876.json,informatiker/Qwen2-7B-Instruct-abliterated,False,7615616512,0.9383132774609197,0.7595159044446702,0.7494506483896471,0.7805289337434647 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/informatiker/Qwen2-7B-Instruct-abliterated/results_2024-09-19T01-44-47.571825.json,informatiker/Qwen2-7B-Instruct-abliterated,False,7615616512,0.9383132774609197,0.7595159044446702,0.7494506483896471,0.7805289337434647 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Intel/neural-chat-7b-v3-1/results_2024-02-25T06-21-33.008420.json,Intel/neural-chat-7b-v3-1,False,7241732096,0.9268770228292367,0.7658477385894799,0.7840135895978708,0.8905574366528357 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Intel/neural-chat-7b-v3-3/results_2024-02-21T22-54-50.520595.json,Intel/neural-chat-7b-v3-3,False,7241732096,0.9140545431322211,0.7587721518241414,0.7147222222222223,0.8653967318817455 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm-20b/results_2024-02-07T08-37-33.747354.json,internlm/internlm-20b,False,20088714240,0.8925812286460775,0.6923784013644456,0.7342526326652933,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm-20b/results_2024-02-17T06-45-40.211098.json,internlm/internlm-20b,False,20088714240,0.8918518746598784,0.8147012269363251,0.774612205615942,0.8009141994772555 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm-7b/results_2024-02-07T07-19-31.042447.json,internlm/internlm-7b,False,7321948160,0.4549404713050996,0.08878825013708527,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm-7b/results_2024-02-17T06-45-37.000547.json,internlm/internlm-7b,False,7321948160,0.6409999089185318,0.32692548050482767,0.4396551724137931,0.6350899298130418 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-1_8b/results_2024-02-23T18-43-57.812118.json,internlm/internlm2-1_8b,False,1889110016,0.670446278362269,0.20280834503832296,0.5658898379556399,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-1_8b/results_2024-04-21T04-02-08.292884.json,internlm/internlm2-1_8b,False,1889110016,0.6656195653662345,0.21185312761987277,0.5777374704126788,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-1_8b/results_2024-04-21T14-43-54.670161.json,internlm/internlm2-1_8b,False,1889110016,0.723384521145955,0.22468395894427515,0.46684697067104125,0.39991043439319296 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-20b/results_2024-02-17T16-58-30.727692.json,internlm/internlm2-20b,False,19861149696,0.9015420190554897,0.8256445175880561,0.6454202534094207,0.7593890574327858 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-20b/results_2024-04-21T04-22-51.086507.json,internlm/internlm2-20b,False,19861149696,0.9060451872187814,0.7952162262919344,0.5841656548248535,0.8660229181072805 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-20b/results_2024-04-21T07-00-47.342945.json,internlm/internlm2-20b,False,19861149696,0.9015420190554897,0.8256445175880561,0.6454202534094207,0.7593890574327858 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-7b/results_2024-02-17T15-30-06.358225.json,internlm/internlm2-7b,False,7737708544,0.8758216175359033,0.7912849899676124,0.4617741808753045,0.6885926784506906 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-7b/results_2024-04-21T04-00-10.218229.json,internlm/internlm2-7b,False,7737708544,0.8733375622238231,0.7913086350010642,0.4396551724137931,0.689607719224915 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-7b/results_2024-04-21T14-47-30.164842.json,internlm/internlm2-7b,False,7737708544,0.882009880708519,0.8114788763197688,0.6006743531496006,0.6798394312598386 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-base-20b/results_2024-02-17T18-21-48.308894.json,internlm/internlm2-base-20b,False,19861149696,0.8027253755556476,0.7829855914043001,0.6753571071820996,0.6954053679859175 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-base-20b/results_2024-04-22T13-51-39.728363.json,internlm/internlm2-base-20b,False,19861149696,0.9113483835168366,0.8221230862082967,0.5596523343471179,0.7213545518615516 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-base-20b/results_2024-04-22T16-19-07.823561.json,internlm/internlm2-base-20b,False,19861149696,0.8027253755556476,0.7832161110060334,0.6753571071820996,0.6954053679859175 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-base-7b/results_2024-02-17T15-56-34.059613.json,internlm/internlm2-base-7b,False,7737708544,0.7271807845224132,0.6816429162037002,0.4396551724137931,0.5197536000233555 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-base-7b/results_2024-04-22T13-59-41.726405.json,internlm/internlm2-base-7b,False,7737708544,0.7384782301051936,0.6846141261557446,0.4396551724137931,0.5114405658342088 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-base-7b/results_2024-04-22T14-00-18.032638.json,internlm/internlm2-base-7b,False,7737708544,0.7882519462137296,0.7455395787744089,0.4956660533374715,0.634179457007547 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-1_8b/results_2024-02-23T19-22-03.349614.json,internlm/internlm2-chat-1_8b,False,1889110016,0.7564501526844221,0.5253914094373616,0.6605087532643389,0.5865292253874497 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-1_8b/results_2024-04-20T01-32-50.384447.json,internlm/internlm2-chat-1_8b,False,1889110016,0.7640453149702802,0.5334218649148808,0.5935038177190644,0.49036887586128136 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-1_8b/results_2024-04-21T04-06-52.514986.json,internlm/internlm2-chat-1_8b,False,1889110016,0.7568367402860745,0.5255865556270367,0.6605087532643389,0.585505071021847 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-1_8b-sft/results_2024-04-21T14-44-58.899037.json,internlm/internlm2-chat-1_8b-sft,False,1889110016,0.7779868068914876,0.547937612914106,0.6225551060103202,0.525101763907734 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-1_8b-sft/results_2024-04-21T14-46-03.038693.json,internlm/internlm2-chat-1_8b-sft,False,1889110016,0.7779734389903882,0.5237417363095832,0.6646911146011257,0.6171543767345573 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-20b/results_2024-02-24T08-23-06.516805.json,internlm/internlm2-chat-20b,False,19861149696,0.8938164607999035,0.8005865862472271,0.808378651654688,0.6808266922301873 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-20b/results_2024-04-19T18-41-41.782142.json,internlm/internlm2-chat-20b,False,19861149696,0.9121640418959764,0.8274756528402901,0.788058906760876,0.798735223073727 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-20b/results_2024-04-21T14-44-01.464293.json,internlm/internlm2-chat-20b,False,19861149696,0.8938164607999035,0.8005865862472271,0.808378651654688,0.6808266922301873 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-20b-sft/results_2024-04-21T22-19-07.591394.json,internlm/internlm2-chat-20b-sft,False,19861149696,0.9138008461911635,0.8328354997872777,0.8039609932643008,0.7946698574084508 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-20b-sft/results_2024-04-22T06-09-22.928789.json,internlm/internlm2-chat-20b-sft,False,19861149696,0.8909513995593468,0.8010479909085628,0.56880564255592,0.6817472664224469 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-7b/results_2024-02-24T03-53-59.489842.json,internlm/internlm2-chat-7b,False,7737708544,0.9101042574287174,0.8254942285584971,0.7825802417067815,0.8812385531135531 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-7b/results_2024-04-19T22-46-45.317633.json,internlm/internlm2-chat-7b,False,7737708544,0.9096027374760287,0.8155847966552392,0.7206513737654834,0.7927824467627244 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-7b/results_2024-04-21T17-46-45.386513.json,internlm/internlm2-chat-7b,False,7737708544,0.9109214914520927,0.8238554887894789,0.7855415483791437,0.8725908838061582 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-7b-sft/results_2024-04-21T22-34-09.862191.json,internlm/internlm2-chat-7b-sft,False,7737708544,0.9100613218260277,0.8269099377672186,0.816261444168421,0.8842741416998843 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-chat-7b-sft/results_2024-04-22T13-58-01.674305.json,internlm/internlm2-chat-7b-sft,False,7737708544,0.9021140593925561,0.8193710267668627,0.7312635184242251,0.7759525197392825 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2-wqx-20b/results_2024-06-12T06-17-43.376516.json,internlm/internlm2-wqx-20b,False,19861149696,0.7529503037030962,0.7810029697049212,0.4396551724137931,0.40246064045568575 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2_5-1_8b/results_2024-08-11T08-17-24.270891.json,internlm/internlm2_5-1_8b,False,1889110016,0.4014150608539269,0.35026225315542026,0.4386873920552677,0.3594117505398441 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2_5-1_8b-chat/results_2024-08-11T07-25-10.789043.json,internlm/internlm2_5-1_8b-chat,False,1889110016,0.7358951629011237,0.4107341119021913,0.4396551724137931,0.6161427533976553 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2_5-20b/results_2024-08-11T08-47-37.245146.json,internlm/internlm2_5-20b,False,19861149696,0.9121707257446139,0.7695155175519195,0.7302015101892285,0.8932969949753422 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2_5-20b-chat/results_2024-08-11T09-06-11.616640.json,internlm/internlm2_5-20b-chat,False,19861149696,0.5912325183134722,0.8220632161489191,0.4284452296819788,0.46820156871019825 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2_5-7b/results_2024-07-05T17-44-03.473493.json,internlm/internlm2_5-7b,False,7737708544,0.8190306959836143,0.5902389147315774,0.4396551724137931,0.35379255536914095 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2_5-7b-chat/results_2024-07-06T01-17-51.647558.json,internlm/internlm2_5-7b-chat,False,7737708544,0.8998640109993313,0.799391039043716,0.7338973799126638,0.8648841174497106 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/internlm/internlm2_5-7b-chat-1m/results_2024-07-06T01-31-37.378717.json,internlm/internlm2_5-7b-chat-1m,False,7737708544,0.910065346903687,0.8151959738123824,0.7336944501850905,0.8238618977870904 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/invalid-coder/Sakura-SOLAR-Instruct-CarbonVillain-en-10.7B-v2-slerp/results_2024-05-19T12-51-48.449854.json,invalid-coder/Sakura-SOLAR-Instruct-CarbonVillain-en-10.7B-v2-slerp,False,10731524096,0.9293286384185251,0.8165192082173185,0.756614128081815,0.7869372225745086 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Isaak-Carter/Josiefied-Qwen2.5-7B-Instruct-abliterated/results_2024-09-24T02-32-40.305523.json,Isaak-Carter/Josiefied-Qwen2.5-7B-Instruct-abliterated,False,7615616512,0.9383152538950544,0.7682557508318288,0.8300209205020921,0.7830583484192929 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Isaak-Carter/Josiefied-Qwen2.5-7B-Instruct-abliterated-v2/results_2024-09-24T01-37-19.261907.json,Isaak-Carter/Josiefied-Qwen2.5-7B-Instruct-abliterated-v2,False,7615616512,0.9403575264313373,0.7765350909675457,0.8537571576602413,0.7395348837209302 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/J-AI/Mistral_7b_ptbr_icatalogo/results_2024-05-28T09-50-52.381944.json,J-AI/Mistral_7b_ptbr_icatalogo,False,7241732096,0.8908588572785072,0.4870714404135913,0.7366085211485026,0.7774362778867046 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/J-LAB/BRisa-7B-Instruct-v0.2/results_2024-04-19T10-58-40.659391.json,J-LAB/BRisa-7B-Instruct-v0.2,False,7241732096,0.914959114959115,0.7360504820365534,0.6830556684274685,0.7427748086927932 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/J-LAB/BRisa-7B-Instruct-v0.2/results_2024-04-19T16-12-48.574869.json,J-LAB/BRisa-7B-Instruct-v0.2,False,7241732096,0.9133356979253711,0.7369091629413509,0.6808719560094265,0.7400764917752776 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/J-LAB/Brisa_v0.2_sft/results_2024-10-26T01-39-21.524545.json,J-LAB/Brisa_v0.2_sft,False,7241748480,0.8400657715367597,0.5851596486591,0.4830369649343558,0.6671689268374351 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/J-LAB/FluxiIA-Small/results_2024-11-19T08-26-18.915800.json,J-LAB/FluxiIA-Small,False,7241748480,0.9178021640979843,0.7182073154044585,0.4396551724137931,0.7477510136352553 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/J-LAB/FluxiIA-Small_Brisa/results_2024-11-19T01-35-25.620513.json,J-LAB/FluxiIA-Small_Brisa,False,7241748480,0.883476634386219,0.5204016200069466,0.6144573702739936,0.7702344358104989 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/J-LAB/Nemo_RolePlay_PTBR_new/results_2024-09-11T01-40-23.266529.json,J-LAB/Nemo_RolePlay_PTBR_new,False,12247782400,0.844887648522527,0.6967016218317799,0.6821792656298142,0.701855513576498 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/J-LAB/Qwen2.5_Thinik_base_v1.0/results_2024-10-20T02-55-07.871989.json,J-LAB/Qwen2.5_Thinik_base_v1.0,False,3085938688,0.0,0.0,0.17721518987341772,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/J-LAB/Thinik/results_2024-10-15T05-47-18.480390.json,J-LAB/Thinik,False,3085938688,0.0,0.0,0.17721518987341772,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/jaspionjader/f-9-8b/results_2025-01-31T01-30-58.169599.json,jaspionjader/f-9-8b,False,8030261248,0.9264468261832206,0.7640541728535835,0.695618394865955,0.8505719522733384 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/jeonsworld/CarbonVillain-en-10.7B-v4/results_2024-05-07T13-32-38.659974.json,jeonsworld/CarbonVillain-en-10.7B-v4,False,10731524096,0.9305548139050568,0.8159105813521528,0.7604170323305367,0.7901253336035945 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/JJhooww/Mistral-7B-v0.2-Base_ptbr/results_2024-04-13T17-20-04.237102.json,JJhooww/Mistral-7B-v0.2-Base_ptbr,False,7241732096,0.9011456831413249,0.7251095355270992,0.6904462094795298,0.7961717229751414 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/JJhooww/Mistral-7B-v0.2-Base_ptbr/results_2024-04-20T03-39-36.902005.json,JJhooww/Mistral-7B-v0.2-Base_ptbr,False,7241732096,0.9019128003131698,0.16704630760888603,0.566892243623113,0.7250569715248458 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/JJhooww/MistralReloadBR_v2_ptbr/results_2024-03-09T04-58-37.486266.json,JJhooww/MistralReloadBR_v2_ptbr,False,7241732096,0.9101172201226876,0.745635698648774,0.4760412001791312,0.7982678280152018 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/JJhooww/Mistral_Reload_v4_Functions/results_2024-06-16T00-09-18.777452.json,JJhooww/Mistral_Reload_v4_Functions,False,7241748480,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/JJhooww/Mistral_Relora_Step2k/results_2024-03-09T08-42-21.029909.json,JJhooww/Mistral_Relora_Step2k,False,7241732096,0.9113496854193482,0.7074610038971542,0.6526577185427341,0.8133973664850924 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/JJhooww/Mistral_Relora_Step2k/results_2024-04-20T03-09-26.234801.json,JJhooww/Mistral_Relora_Step2k,False,7241732096,0.9121625173669783,0.7065946896645577,0.6466313961043266,0.8143254279726638 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/JJhooww/Qwen2-7B_reload_v1/results_2024-06-17T01-31-38.104074.json,JJhooww/Qwen2-7B_reload_v1,False,7615616512,0.939940066242574,0.7603163745995428,0.7801986894175493,0.7412705678079557 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/jondurbin/bagel-8b-v1.0/results_2024-05-20T22-30-54.519411.json,jondurbin/bagel-8b-v1.0,False,8030261248,0.9263427109974425,0.7317294420374749,0.6997642711518859,0.8530957901862226 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/jondurbin/bagel-dpo-34b-v0.2/results_2024-05-19T18-00-27.350171.json,jondurbin/bagel-dpo-34b-v0.2,False,34388917248,0.915343858772782,0.7892772408353919,0.8386087984134318,0.7278548095635251 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/jondurbin/bagel-dpo-34b-v0.5/results_2024-06-16T13-20-54.422898.json,jondurbin/bagel-dpo-34b-v0.5,False,34388917248,0.911645281793081,0.7746321085040028,0.6539479482692303,0.8726822177956144 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Joseph717171/Llama-3.1-SuperNova-8B-Lite_TIES_with_Base/results_2024-10-05T08-11-16.769925.json,Joseph717171/Llama-3.1-SuperNova-8B-Lite_TIES_with_Base,False,8030261248,0.9264609674361474,0.7495427536661012,0.7135850719390003,0.8683608058608059 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/JosephusCheung/LL7M/results_2024-04-23T00-09-23.964862.json,JosephusCheung/LL7M,False,6799233024,0.656772420167965,0.21905517553818948,0.5111651047090131,0.7436810107109835 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/josu/gpt-neo-pt-1.3B/results_2024-02-07T23-30-57.977192.json,josu/gpt-neo-pt-1.3B,False,1315575808,0.10905681673372047,,0.0026041666666666665,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/josu/gpt-neo-pt-1.3B/results_2024-03-08T06-03-42.405393.json,josu/gpt-neo-pt-1.3B,False,1315575808,0.005913692609064442,0.03783913229548836,0.0,0.24126984126984127 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/josu/gpt-neo-pt-br/results_2024-02-07T22-44-00.009771.json,josu/gpt-neo-pt-br,False,125198592,0.09577034137775386,0.006121287694326073,0.16583992675494338,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/josu/gpt-neo-pt-br/results_2024-03-08T07-02-36.204782.json,josu/gpt-neo-pt-br,False,125198592,0.14275161246863377,0.015448479902093387,0.25540986008351035,0.22092646582442502 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/jpacifico/Chocolatine-14B-Instruct-4k-DPO/results_2024-09-01T04-36-58.986732.json,jpacifico/Chocolatine-14B-Instruct-4k-DPO,False,13960238080,0.9309026075700955,0.7151811232953084,0.6798499329309189,0.8574935502655591 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/jsfs11/MixtureofMerges-MoE-4x7b-v4/results_2024-05-19T03-46-55.771244.json,jsfs11/MixtureofMerges-MoE-4x7b-v4,False,24153690112,0.9223851546589933,0.7765615051221151,0.7848580929549145,0.8337715452765142 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/jsfs11/MixtureofMerges-MoE-4x7b-v5/results_2024-06-29T04-47-16.492324.json,jsfs11/MixtureofMerges-MoE-4x7b-v5,False,24153690112,0.9236107924352364,0.7773330335553215,0.7823323438856359,0.8094954980227307 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/kaist-ai/janus-7b/results_2024-10-12T07-47-24.575443.json,kaist-ai/janus-7b,False,7241732096,0.8830322130536242,0.7563225505379452,0.7238984901302048,0.7740471749282323 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/kaist-ai/janus-dpo-7b/results_2024-10-12T01-43-03.009383.json,kaist-ai/janus-dpo-7b,False,7241732096,0.8835390779576071,0.7555679428752237,0.7147180097249974,0.8285951442207926 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/kaist-ai/janus-orpo-7b/results_2024-10-12T02-52-36.378828.json,kaist-ai/janus-orpo-7b,False,7241732096,0.8929635682014804,0.7152807590839677,0.7102054064079381,0.7251652280474239 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/kaist-ai/janus-rm-7b/results_2024-10-18T05-29-16.806255.json,kaist-ai/janus-rm-7b,False,7241732096,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/kaist-ai/mistral-orpo-capybara-7k/results_2024-04-30T14-53-05.191999.json,kaist-ai/mistral-orpo-capybara-7k,False,7241732096,0.8769269622131259,0.6990866776197457,0.604143159962043,0.670893578525416 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/kaitchup/Qwen1.5-7B-bnb-4bit/results_2024-04-25T00-36-40.084562.json,kaitchup/Qwen1.5-7B-bnb-4bit,False,7721324544,0.9019605225568226,0.6550922155680559,0.446593203434771,0.8170420494583293 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/kekmodel/StopCarbon-10.7B-v5/results_2024-05-21T02-58-06.102309.json,kekmodel/StopCarbon-10.7B-v5,False,10731524096,0.9293286384185251,0.8169705299837213,0.7634310134310134,0.7870538006037701 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Kquant03/CognitiveFusion2-4x7B-BF16/results_2024-05-19T01-32-18.922295.json,Kquant03/CognitiveFusion2-4x7B-BF16,False,24153690112,0.9199303114955995,0.7779218286784337,0.7730900759529709,0.8170741905271608 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Kukedlc/NeuralExperiment-7b-MagicCoder-v7.5/results_2024-07-28T01-32-22.165106.json,Kukedlc/NeuralExperiment-7b-MagicCoder-v7.5,False,7241732096,0.9207426934587088,0.7806469957644414,0.7641446815289443,0.8412698412698412 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Kukedlc/NeuralLLaMa-3-8b-DT-v0.1/results_2024-05-20T00-13-17.559065.json,Kukedlc/NeuralLLaMa-3-8b-DT-v0.1,False,8030261248,0.9186925298411294,0.7743511836631678,0.74972872694179,0.850054890620398 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Kukedlc/NeuralLLaMa-3-8b-ORPO-v0.3/results_2024-05-29T10-29-45.644965.json,Kukedlc/NeuralLLaMa-3-8b-ORPO-v0.3,False,8030277632,0.8174821631540747,0.6525382106815406,0.5677269590313069,0.5042722896900733 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Kukedlc/NeuralLLaMa-3-8b-ORPO-v0.4/results_2024-06-15T16-20-59.451134.json,Kukedlc/NeuralLLaMa-3-8b-ORPO-v0.4,False,8030277632,0.8574769945658907,0.7331740818932964,0.6731697506033789,0.7806621111874374 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Kukedlc/NeuralSynthesis-7B-v0.1/results_2024-06-15T14-00-57.278364.json,Kukedlc/NeuralSynthesis-7B-v0.1,False,7241732096,0.924014535978148,0.7798626544905576,0.7806541284802154,0.8201619169361105 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Kukedlc/NeuralSynthesis-7B-v0.3/results_2024-08-07T06-51-04.901695.json,Kukedlc/NeuralSynthesis-7B-v0.3,False,7241732096,0.922380439977678,0.7790400956364452,0.7796173768257517,0.8186191231446596 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Kukedlc/NeuralSynthesis-7b-v0.4-slerp/results_2024-06-16T02-26-33.125166.json,Kukedlc/NeuralSynthesis-7b-v0.4-slerp,False,7241732096,0.9227873017727286,0.7796159438413869,0.7812802842683321,0.8163838486760461 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Lambent/qwen2.5-reinstruct-alternate-lumen-14B/results_2024-10-03T01-46-42.973832.json,Lambent/qwen2.5-reinstruct-alternate-lumen-14B,False,14765947904,0.9468953893904251,0.8356280833811742,0.8031640750181548,0.8865641025641027 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/langtech-languagemodeling/IberianLLM-7B-Instruct/results_2025-07-15T03-53-32.324544.json,langtech-languagemodeling/IberianLLM-7B-Instruct,False,7768117248,0.8692754613807245,0.5275807250702291,0.5416854693140793,0.7042961328675614 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/leafspark/Llama-3.1-8B-MultiReflection-Instruct/results_2024-09-20T01-40-40.081160.json,leafspark/Llama-3.1-8B-MultiReflection-Instruct,False,8030261248,0.9207007677588055,0.7716200183368191,0.674173738017691,0.8755151098901099 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/leliuga/Phi-3-mini-4k-instruct-bnb-4bit/results_2024-04-25T00-35-53.556686.json,leliuga/Phi-3-mini-4k-instruct-bnb-4bit,False,3821079552,0.8928594338008895,0.7075972430571313,0.6790123456790124,0.7930671758683527 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lemon-mint/gemma-ko-7b-instruct-v0.62/results_2024-05-20T04-28-34.823546.json,lemon-mint/gemma-ko-7b-instruct-v0.62,False,8537680896,0.9313226285446337,0.757678890960136,0.8025691008147149,0.8976498838719835 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/results_2024-08-11T07-21-08.591192.json,LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct,False,7818448896,0.872773896630295,0.7643433566585385,0.7424139492753623,0.7379657769346437 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lightblue/suzume-llama-3-8B-multilingual/results_2024-05-17T23-12-15.124105.json,lightblue/suzume-llama-3-8B-multilingual,False,8030261248,0.9071794079450122,0.7527864809827423,0.6979047619047619,0.8267028999321541 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lightblue/suzume-llama-3-8B-multilingual-orpo-borda-full/results_2024-06-14T13-03-18.875570.json,lightblue/suzume-llama-3-8B-multilingual-orpo-borda-full,False,8030261248,0.9109471774508449,0.7282984602160687,0.7263950844047932,0.8338575591714755 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lightblue/suzume-llama-3-8B-multilingual-orpo-borda-half/results_2024-06-16T07-20-37.794606.json,lightblue/suzume-llama-3-8B-multilingual-orpo-borda-half,False,8030261248,0.9202903451229684,0.7413907246289755,0.6948528721386589,0.8274347536420472 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lightblue/suzume-llama-3-8B-multilingual-orpo-borda-half/results_2024-06-18T13-08-46.981373.json,lightblue/suzume-llama-3-8B-multilingual-orpo-borda-half,False,8030261248,0.9202903451229684,0.7413907246289755,0.6948528721386589,0.8274347536420472 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lightblue/suzume-llama-3-8B-multilingual-orpo-borda-top25/results_2024-06-13T00-28-44.151475.json,lightblue/suzume-llama-3-8B-multilingual-orpo-borda-top25,False,8030261248,0.9187052298614165,0.7544804305824591,0.6932700620255123,0.82846557682084 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lightblue/suzume-llama-3-8B-multilingual-orpo-borda-top75/results_2024-06-14T13-04-49.415893.json,lightblue/suzume-llama-3-8B-multilingual-orpo-borda-top75,False,8030261248,0.9071003308602819,0.7074059085676956,0.5843826237506886,0.8407244760972945 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/liminerity/M7-7b/results_2024-07-13T18-52-55.048115.json,liminerity/M7-7b,False,7241732096,0.9219689860564255,0.7788708607433608,0.7785648031031285,0.8195498766124012 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/LiteAI/Hare-1.1B-base/results_2024-07-25T09-16-21.440502.json,LiteAI/Hare-1.1B-base,False,1123117056,0.335683441456502,0.00629547079812791,0.4396551724137931,0.33834570221775717 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/LiteAI/Hare-1.1B-Chat/results_2024-07-25T08-38-24.361382.json,LiteAI/Hare-1.1B-Chat,False,1123641344,0.3436190894130554,0.028209892267941836,0.4396551724137931,0.37450625769533197 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lmsys/vicuna-13b-v1.5/results_2024-03-08T12-42-13.451104.json,lmsys/vicuna-13b-v1.5,False,13015864320,0.8705016738481687,0.7373103005646825,0.7042248591130246,0.7602114237326661 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lmsys/vicuna-33b-v1.3/results_2024-12-07T12-25-45.401959.json,lmsys/vicuna-33b-v1.3,False,32528943616,0.822100949248243,0.740526946373418,0.7265413060445252,0.7578945795286416 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lmsys/vicuna-7b-v1.5/results_2024-03-08T09-21-18.863920.json,lmsys/vicuna-7b-v1.5,False,6738415616,0.7957441266193219,0.507613025680697,0.59375,0.6980836377464594 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lrds-code/boana-7b-instruct/results_2024-02-06T17-34-10.551104.json,lrds-code/boana-7b-instruct,False,6738415616,0.3582588385476761,0.31776150050082763,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lrds-code/boana-7b-instruct/results_2024-02-18T04-08-18.900803.json,lrds-code/boana-7b-instruct,False,6738415616,0.4883903675731297,0.3755961900599195,0.4396551724137931,0.8499559054089365 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lrds-code/samba-1.1B/results_2024-02-06T17-34-10.550582.json,lrds-code/samba-1.1B,False,1100048384,0.45131674757281554,0.10316280984073979,0.17721518987341772,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lrds-code/samba-1.1B/results_2024-02-18T01-47-54.897005.json,lrds-code/samba-1.1B,False,1100048384,0.3333333333333333,0.013001228192609664,0.17721518987341772,0.5368004230609069 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lucianosb/boto-27B/results_2024-07-13T01-31-41.225952.json,lucianosb/boto-27B,False,27227128320,0.5150650140137017,0.18721992880494653,0.5014563435616067,0.5417144397961452 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lucianosb/boto-7B/results_2024-04-12T06-05-03.197673.json,lucianosb/boto-7B,False,7241732096,0.8958312300388664,0.6987313557551754,0.5556747433955914,0.7703575597141548 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lucianosb/boto-7B-v1.1/results_2024-04-23T06-44-43.435304.json,lucianosb/boto-7B-v1.1,False,7241732096,0.8812212765701137,0.6371302436426871,0.48833520895165566,0.7597439961236006 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lucianosb/boto-7B-v1.2/results_2024-05-24T00-45-26.128118.json,lucianosb/boto-7B-v1.2,False,7248023552,0.9072699884508002,0.7236519452396313,0.7237856993954555,0.7457919897159844 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lucianosb/boto-7B-v1.2/results_2024-05-25T07-06-27.613526.json,lucianosb/boto-7B-v1.2,False,7248023552,0.9056315701554467,0.7239089091760973,0.7078225796654234,0.7452877427374001 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lucianosb/boto-9B/results_2024-07-11T01-31-45.512376.json,lucianosb/boto-9B,False,9241705984,0.8938265179527836,0.7658585049770954,0.5686096767405345,0.7788193164463637 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/lucianosb/boto-9B-it/results_2024-07-13T04-40-47.573532.json,lucianosb/boto-9B-it,False,9241705984,0.9403593178887337,0.8082771327645331,0.7759016051698978,0.8285378657706858 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/M4-ai/tau-0.5B/results_2024-04-23T00-49-38.450870.json,M4-ai/tau-0.5B,False,463987712,0.39285662181494563,0.07216057977107923,0.4396551724137931,0.33301572177227245 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/M4-ai/tau-0.5B-instruct-DPOP/results_2024-04-23T02-24-34.579672.json,M4-ai/tau-0.5B-instruct-DPOP,False,463987712,0.6250683495142939,0.14960279813437427,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/M4-ai/tau-1.8B/results_2024-04-23T02-51-19.597157.json,M4-ai/tau-1.8B,False,1836828672,0.6240877656394997,0.19269473597203718,0.5980814057737135,0.41976405672054573 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/macadeliccc/Samantha-Qwen-2-7B/results_2024-07-17T05-17-28.591280.json,macadeliccc/Samantha-Qwen-2-7B,False,7615616512,0.9267866138357781,0.7998325586024508,0.7864519408388193,0.8285951442207926 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Magpie-Align/Llama-3-8B-Magpie-Align-SFT-v0.3/results_2024-08-12T03-10-50.370471.json,Magpie-Align/Llama-3-8B-Magpie-Align-SFT-v0.3,False,8030261248,0.8818500830491456,0.7136318187555281,0.6780660377358491,0.8563282210891179 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Magpie-Align/Llama-3-8B-Magpie-Align-v0.3/results_2024-08-08T06-53-40.680854.json,Magpie-Align/Llama-3-8B-Magpie-Align-v0.3,False,8030261248,0.8735121156790977,0.616645726427636,0.6134815272203353,0.8245849679145775 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Magpie-Align/Llama-3.1-8B-Magpie-Align-SFT-v0.1/results_2024-08-12T01-51-52.471144.json,Magpie-Align/Llama-3.1-8B-Magpie-Align-SFT-v0.1,False,8030285824,0.8898593896790856,0.7501875514409228,0.6880499120140777,0.8393926345935507 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Magpie-Align/MagpieLM-8B-Chat-v0.1/results_2024-10-04T05-53-00.836549.json,Magpie-Align/MagpieLM-8B-Chat-v0.1,False,8030261248,0.8981976332283619,0.11974003365018737,0.5374826028332357,0.7909906653455931 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Magpie-Align/MagpieLM-8B-Chat-v0.1/results_2025-05-07T15-22-16.111876.json,Magpie-Align/MagpieLM-8B-Chat-v0.1,False,0,0.888146118500855,0.3012536440181936,0.5342091520965496,0.7705044730987287 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Magpie-Align/MagpieLM-8B-SFT-v0.1/results_2024-10-04T07-18-43.453507.json,Magpie-Align/MagpieLM-8B-SFT-v0.1,False,8030261248,0.8752699434198592,0.7220511644434541,0.7227240790926428,0.8505056727758098 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Magpie-Align/MagpieLM-8B-SFT-v0.1/results_2025-05-07T15-05-27.217601.json,Magpie-Align/MagpieLM-8B-SFT-v0.1,False,0,0.8394850106817857,0.6798371427313809,0.603366491664364,0.7847849078879405 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MagusCorp/grpo_lora_enem_llama3_7b/results_2025-02-12T16-40-51.376680.json,MagusCorp/grpo_lora_enem_llama3_7b,False,8030261248,0.9256174008161892,0.7947537951592014,0.7002788356171814,0.8619417320149316 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MagusCorp/legislinho/results_2024-04-13T08-30-14.215121.json,MagusCorp/legislinho,False,7241748480,0.8870184075342467,0.6776356777228696,0.6379609737375439,0.7264113460475401 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/maldv/Qwentile2.5-32B-Instruct/results_2024-12-22T12-03-47.113605.json,maldv/Qwentile2.5-32B-Instruct,False,32763876352,0.9287015115643781,0.8374290907001216,0.7905927835051547,0.9249953696121036 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MarinaraSpaghetti/NemoReRemix-12B/results_2024-09-21T02-41-01.308995.json,MarinaraSpaghetti/NemoReRemix-12B,False,12247782400,0.9160102526939263,0.8022910212385954,0.7826145905832798,0.8195498766124012 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/maritaca-ai/sabia-7b/results_2024-02-06T04-39-04.874891.json,maritaca-ai/sabia-7b,False,6738415616,0.4423440331551014,0.15555684597410213,0.5619638242894056,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/maritaca-ai/sabia-7b/results_2024-02-16T23-09-41.630193.json,maritaca-ai/sabia-7b,False,6738415616,0.4668233911254775,0.018856147318320138,0.5833926791055405,0.619251346675534 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/maritaca-ai/sabia-7b/results_2024-04-19T18-19-54.776223.json,maritaca-ai/sabia-7b,False,6738415616,0.45246586529977384,0.019944161517541175,0.5925653394380399,0.6112101828964611 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/matheusrdgsf/cesar-ptbr/results_2024-05-27T22-31-08.120812.json,matheusrdgsf/cesar-ptbr,False,1143017472,0.8747602497221476,0.6849438767598355,0.7380971253434792,0.8330124462550406 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/matsuo-lab/weblab-10b/results_2024-02-15T00-41-06.033489.json,matsuo-lab/weblab-10b,False,10712113664,0.33424046164366045,0.07121886203158499,0.5263362445886277,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/matsuo-lab/weblab-10b/results_2024-04-05T12-07-37.917373.json,matsuo-lab/weblab-10b,False,10712113664,0.5531806064965059,0.05039357461980308,0.5322038013318535,0.3349186726374015 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/maywell/Qwen2-7B-Multilingual-RP/results_2024-07-26T03-26-28.841869.json,maywell/Qwen2-7B-Multilingual-RP,False,7615616512,0.9154293080465958,0.750982061619474,0.770305276407592,0.655900445157104 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/maywell/Synatra-7B-v0.3-RP/results_2024-05-18T21-35-16.166862.json,maywell/Synatra-7B-v0.3-RP,False,7241748480,0.8982291377095245,0.6435083220682054,0.3857639628131432,0.7439554115565462 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MaziyarPanahi/calme-2.1-qwen2-7b/results_2025-02-04T12-12-25.697301.json,MaziyarPanahi/calme-2.1-qwen2-7b,False,7615616512,0.927948321222374,0.774196296621951,0.7642364889372506,0.7514438502673797 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MaziyarPanahi/calme-2.2-qwen2-7b/results_2025-02-04T13-14-47.416975.json,MaziyarPanahi/calme-2.2-qwen2-7b,False,7615616512,0.9088742688391502,0.768339263906459,0.8198022445144928,0.8005158927228114 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MaziyarPanahi/calme-2.3-qwen2-7b/results_2025-02-04T14-24-00.630908.json,MaziyarPanahi/calme-2.3-qwen2-7b,False,7615616512,0.8757778341603875,0.7701409739179119,0.7313748190235345,0.8160029630691661 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MaziyarPanahi/calme-2.4-qwen2-7b/results_2025-02-04T15-24-00.944476.json,MaziyarPanahi/calme-2.4-qwen2-7b,False,7615616512,0.9246555544850178,0.7230229756328089,0.6993592017638791,0.7752748274431711 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MaziyarPanahi/calme-2.7-qwen2-7b/results_2024-09-23T01-40-34.661105.json,MaziyarPanahi/calme-2.7-qwen2-7b,False,7615616512,0.9338075797077743,0.8107955502233284,0.6540472994862776,0.7865853658536586 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MaziyarPanahi/Calme-4x7B-MoE-v0.1/results_2024-06-16T23-38-24.109054.json,MaziyarPanahi/Calme-4x7B-MoE-v0.1,False,24153690112,0.9207511579149721,0.7871041906910172,0.7526873404995746,0.8025288840687826 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MaziyarPanahi/Calme-4x7B-MoE-v0.2/results_2024-07-13T10-03-53.912535.json,MaziyarPanahi/Calme-4x7B-MoE-v0.2,False,24153690112,0.9195254857821783,0.7862069927082468,0.7532485313153441,0.8110268392962026 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MaziyarPanahi/Llama-3-8B-Instruct-v0.10/results_2024-08-13T22-35-47.080828.json,MaziyarPanahi/Llama-3-8B-Instruct-v0.10,False,8030261248,0.9162158773450131,0.7289740656496875,0.7511456136837518,0.8618806417693999 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MaziyarPanahi/Llama-3-8B-Instruct-v0.8/results_2024-06-12T10-17-20.117216.json,MaziyarPanahi/Llama-3-8B-Instruct-v0.8,False,8030261248,0.9182337965638154,0.7259213081034068,0.7473170846635018,0.8684640522875817 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MaziyarPanahi/Llama-3-8B-Instruct-v0.9/results_2024-08-13T23-42-30.931238.json,MaziyarPanahi/Llama-3-8B-Instruct-v0.9,False,8030261248,0.9178565043038492,0.7263342614908861,0.7687727167272447,0.8648344221671547 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MaziyarPanahi/Mistral-7B-Instruct-Aya-101/results_2024-04-17T09-07-30.140283.json,MaziyarPanahi/Mistral-7B-Instruct-Aya-101,False,7241732096,0.9072398971802695,0.7641692139433879,0.6218181818181818,0.8004209608305171 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MaziyarPanahi/Mistral-7B-Instruct-v0.3/results_2024-05-26T08-26-53.282710.json,MaziyarPanahi/Mistral-7B-Instruct-v0.3,False,7248023552,0.9027027027027027,0.7942968956652824,0.6866859299725274,0.8343452944711321 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MaziyarPanahi/Mistral-7B-v0.3/results_2024-05-26T09-27-13.540220.json,MaziyarPanahi/Mistral-7B-v0.3,False,7248023552,0.8790493645404561,0.6748507212873428,0.4777509957775612,0.778367362714571 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MaziyarPanahi/Topxtral-4x7B-v0.1/results_2024-06-13T01-59-37.042342.json,MaziyarPanahi/Topxtral-4x7B-v0.1,False,18516414464,0.9207507876328296,0.7759288997534963,0.7802301835247021,0.8072877590949881 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meraGPT/mera-mix-4x7B/results_2024-06-13T04-11-32.982548.json,meraGPT/mera-mix-4x7B,False,24153690112,0.9215686274509803,0.7819146834745879,0.7733499377334994,0.8551991680384323 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-2-13b-chat-hf/results_2024-04-20T07-13-15.107904.json,meta-llama/Llama-2-13b-chat-hf,False,13015864320,0.7741212250576865,0.584353379958785,0.6814208185925483,0.841311660223144 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-2-13b-hf/results_2024-02-06T00-29-55.201157.json,meta-llama/Llama-2-13b-hf,False,13015864320,0.3686499749065968,0.23670271957356603,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-2-13b-hf/results_2024-02-06T03-05-10.640287.json,meta-llama/Llama-2-13b-hf,False,13015864320,0.3686499749065968,0.23670271957356603,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-2-13b-hf/results_2024-02-16T19-03-44.924653.json,meta-llama/Llama-2-13b-hf,False,13015864320,0.863152475822486,0.5874196845690397,0.4396551724137931,0.8107994907885967 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-2-70b-chat-hf/results_2024-03-05T16-51-51.121987.json,meta-llama/Llama-2-70b-chat-hf,False,68976648192,0.851364454516761,0.718111227933856,0.784438775510204,0.8691622862588227 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-2-70b-hf/results_2024-02-10T08-27-04.991373.json,meta-llama/Llama-2-70b-hf,False,68976648192,0.6275204231005336,0.5673654765601219,0.4760412001791312,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-2-70b-hf/results_2024-03-03T17-30-43.105337.json,meta-llama/Llama-2-70b-hf,False,68976648192,0.8864508287693933,0.7914176878754884,0.45391686964720673,0.8281282837173397 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-2-7b-chat-hf/results_2024-04-20T05-48-24.116596.json,meta-llama/Llama-2-7b-chat-hf,False,6738415616,0.7691790324880676,0.5255824308244532,0.5483679076695784,0.8575226949035188 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-2-7b-hf/results_2024-02-05T23-36-46.051171.json,meta-llama/Llama-2-7b-hf,False,6738415616,0.3632717133913797,0.22243979190632546,0.5421626984126984,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-2-7b-hf/results_2024-02-06T12-34-30.170285.json,meta-llama/Llama-2-7b-hf,False,6738415616,0.3632717133913797,0.22243979190632546,0.5421626984126984,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-2-7b-hf/results_2024-02-16T17-57-01.707450.json,meta-llama/Llama-2-7b-hf,False,6738415616,0.6701769945334302,0.3110321340448155,0.5387420425880172,0.7515535753827598 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-3.1-8B-Instruct/results_2024-10-27T01-39-16.745810.json,meta-llama/Llama-3.1-8B-Instruct,False,8030261248,0.9264373956594324,0.7735762346135125,0.753627075947005,0.8645238289068056 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-3.2-11B-Vision/results_2024-12-05T14-29-14.168846.json,meta-llama/Llama-3.2-11B-Vision,False,9775192080,0.9185916573293528,0.7854934992698208,0.6209988089245935,0.8620361169736637 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-3.2-11B-Vision-Instruct/results_2024-12-05T15-30-34.255935.json,meta-llama/Llama-3.2-11B-Vision-Instruct,False,9775192080,0.9252331082958665,0.7610610599042877,0.7490083195843719,0.8610812746151092 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-3.2-1B/results_2024-12-04T07-26-36.853901.json,meta-llama/Llama-3.2-1B,False,1235814400,0.507334525939177,0.03170374159067819,0.4396551724137931,0.5510511749598876 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-3.2-1B-Instruct/results_2024-12-07T01-41-07.415996.json,meta-llama/Llama-3.2-1B-Instruct,False,1235814400,0.3689023973787093,0.17881741566057854,0.5138369483919222,0.34278702674036976 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-3.2-3B/results_2024-12-04T07-53-29.731609.json,meta-llama/Llama-3.2-3B,False,3212749824,0.8363366289352226,0.46826295540727986,0.4396551724137931,0.7481012827186695 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-3.2-3B-Instruct/results_2024-12-04T08-43-19.289819.json,meta-llama/Llama-3.2-3B-Instruct,False,3212749824,0.850443798601838,0.6189280763466014,0.6696586536440721,0.8630902186306871 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-3.2-90B-Vision-Instruct/results_2024-12-05T19-43-22.283595.json,meta-llama/Llama-3.2-90B-Vision-Instruct,False,87666865192,0.9280721202003339,0.7149057272939108,0.8144399460188934,0.8894129860750637 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Llama-3.3-70B-Instruct/results_2024-12-20T18-22-39.816960.json,meta-llama/Llama-3.3-70B-Instruct,False,70553706496,0.9325977580310909,0.6895446831892964,0.8252071844911102,0.8694271437986395 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Meta-Llama-3-70B/results_2024-05-20T02-14-50.865845.json,meta-llama/Meta-Llama-3-70B,False,70553706496,0.9121480030588014,0.828830729571223,0.7740384615384615,0.8399085191538023 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Meta-Llama-3-70B-Instruct/results_2024-05-20T06-43-19.816797.json,meta-llama/Meta-Llama-3-70B-Instruct,False,70553706496,0.9346307055437728,0.8059930794983008,0.8390001804176516,0.8893657723562216 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Meta-Llama-3-8B/results_2024-04-19T02-51-37.273329.json,meta-llama/Meta-Llama-3-8B,False,8030261248,0.3333333333333333,0.006567397989287458,0.4396551724137931,0.3498440477331034 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Meta-Llama-3-8B/results_2024-04-19T14-09-22.709087.json,meta-llama/Meta-Llama-3-8B,False,8030261248,0.3333333333333333,0.006567397989287458,0.4396551724137931,0.3498440477331034 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Meta-Llama-3-8B/results_2024-04-23T04-30-20.095737.json,meta-llama/Meta-Llama-3-8B,False,8030261248,0.8951198939631281,0.766493362127982,0.4396551724137931,0.7562902556274187 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Meta-Llama-3-8B-Instruct/results_2024-04-19T04-57-44.794348.json,meta-llama/Meta-Llama-3-8B-Instruct,False,8030261248,0.9090728393053975,0.7516552959245816,0.7516225803875942,0.8699139839404014 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Meta-Llama-3.1-70B/results_2024-08-10T11-57-03.700561.json,meta-llama/Meta-Llama-3.1-70B,False,70553706496,0.9151008120892412,0.8291880627497643,0.7727238911455312,0.8228474501117922 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Meta-Llama-3.1-70B-Instruct/results_2024-08-10T06-11-24.432358.json,meta-llama/Meta-Llama-3.1-70B-Instruct,False,70553706496,0.9444173984120974,0.7516561640534767,0.8619961910314953,0.8987376057445633 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Meta-Llama-3.1-70B-Instruct/results_2024-08-12T18-45-00.046141.json,meta-llama/Meta-Llama-3.1-70B-Instruct,False,70553706496,0.9264292893707822,0.71626786467975,0.816949415888985,0.8857359329691872 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Meta-Llama-3.1-8B/results_2024-07-24T05-10-18.781468.json,meta-llama/Meta-Llama-3.1-8B,False,8030261248,0.9028996579924904,0.7700476903056154,0.4396551724137931,0.7648400085532114 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/meta-llama/Meta-Llama-3.1-8B-Instruct/results_2024-07-24T06-11-38.864971.json,meta-llama/Meta-Llama-3.1-8B-Instruct,False,8030261248,0.9226964805305143,0.7459925179142035,0.7263127569280545,0.872010623220419 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/phi-1/results_2024-02-09T03-25-02.286163.json,microsoft/phi-1,False,1418270720,0.3333333333333333,0.010234563981965095,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/phi-1/results_2024-03-08T11-39-34.682713.json,microsoft/phi-1,False,1418270720,0.3447055730752323,0.021169006705963778,0.4396551724137931,0.3469451671487276 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/phi-1_5/results_2024-02-09T04-09-48.434293.json,microsoft/phi-1_5,False,1418270720,0.3333333333333333,0.04512133370254969,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/phi-1_5/results_2024-03-08T11-28-41.393728.json,microsoft/phi-1_5,False,1418270720,0.3333333333333333,0.1301716664085566,0.4396551724137931,0.33301572177227245 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/phi-2/results_2024-02-09T04-17-12.681004.json,microsoft/phi-2,False,2779683840,0.5777512527339992,0.030107129522378703,0.43917169974115616,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/phi-2/results_2024-03-08T09-04-48.667159.json,microsoft/phi-2,False,2779683840,0.38379182258071776,0.08868974435801291,0.43917169974115616,0.5962833412946669 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/Phi-3-medium-128k-instruct/results_2024-05-21T20-04-13.144740.json,microsoft/Phi-3-medium-128k-instruct,False,13960238080,0.9383028991034552,0.7319158064408069,0.7709457029845379,0.798735223073727 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/Phi-3-medium-4k-instruct/results_2024-05-21T17-06-43.203552.json,microsoft/Phi-3-medium-4k-instruct,False,13960238080,0.9354260952140543,0.7031484731508717,0.7501808645324652,0.8307624651156005 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/Phi-3-mini-128k-instruct/results_2024-04-24T06-38-52.920813.json,microsoft/Phi-3-mini-128k-instruct,False,3821079552,0.9039678690992982,0.7212819442471835,0.5566906107709217,0.838141935483871 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/Phi-3-mini-4k-instruct/results_2024-04-24T05-54-26.688190.json,microsoft/Phi-3-mini-4k-instruct,False,3821079552,0.9064497368757004,0.7359927454277331,0.560565346532237,0.8434116670163017 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/Phi-3-small-128k-instruct/results_2024-05-21T19-34-21.860499.json,microsoft/Phi-3-small-128k-instruct,False,7392272384,0.9166091507018962,0.7686357537919025,0.6566507177033494,0.831238953431031 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/Phi-3-small-8k-instruct/results_2024-05-21T16-47-35.100902.json,microsoft/Phi-3-small-8k-instruct,False,7392272384,0.9209267257069647,0.7588287455514673,0.6500897205844656,0.850161239535717 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/Phi-3.5-mini-instruct/results_2024-12-04T05-29-32.708739.json,microsoft/Phi-3.5-mini-instruct,False,3821079552,0.9195132367993755,0.7432365232682981,0.48747910016719864,0.8464153286686453 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/Phi-3.5-MoE-instruct/results_2024-12-07T09-03-45.514107.json,microsoft/Phi-3.5-MoE-instruct,False,41873153344,0.9271660439827878,0.734022501267399,0.7073310983085419,0.83192268443484 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/phi-4/results_2025-01-30T23-02-16.424585.json,microsoft/phi-4,False,14659507200,0.9325614767954681,0.8454012719587444,0.7320140177283034,0.8150854172768274 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/Phi-4-mini-instruct/results_2025-05-04T19-12-11.302575.json,microsoft/Phi-4-mini-instruct,False,3836021760,0.9039996635782954,0.7019460946591871,0.7288654060066742,0.8428542565067522 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/Phi-4-mini-reasoning/results_2025-05-04T20-11-44.128074.json,microsoft/Phi-4-mini-reasoning,False,3836021760,0.5717921140633908,0.012994533178798728,0.3003155439703663,0.4203632778626585 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/Phi-4-reasoning/results_2025-05-06T09-26-57.455990.json,microsoft/Phi-4-reasoning,False,0,0.913226264821906,0.7400535276405864,0.7316881629349694,0.8625362820766541 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/microsoft/Phi-4-reasoning-plus/results_2025-05-06T06-28-51.559439.json,microsoft/Phi-4-reasoning-plus,False,0,0.9212029171770723,0.7514101973166176,0.6975823705605477,0.8293571664475283 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/migtissera/Tess-3-Mistral-Nemo-12B/results_2025-02-05T01-33-37.456610.json,migtissera/Tess-3-Mistral-Nemo-12B,False,12247813120,0.8916342656465571,0.7636855641143607,0.5950841975690101,0.8629541318814027 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/migtissera/Tess-M-v1.3/results_2024-05-26T16-17-28.033562.json,migtissera/Tess-M-v1.3,False,34388945920,0.9145978328635895,0.7832608745111329,0.8054727247159613,0.7397401412080459 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/migtissera/Tess-v2.5-Gemma-2-27B-alpha/results_2024-09-21T01-47-54.487079.json,migtissera/Tess-v2.5-Gemma-2-27B-alpha,False,27227128320,0.9071745703598268,0.6901348190668776,0.6135046611237087,0.8792816023519557 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/migtissera/Tess-v2.5-Gemma-2-27B-alpha/results_2025-07-07T09-11-48.108399.json,migtissera/Tess-v2.5-Gemma-2-27B-alpha,False,27227128320,0.904304391260913,0.6806514407435583,0.5662584393154342,0.8810203985509473 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/migtissera/Tess-v2.5-Phi-3-medium-128k-14B/results_2024-06-23T01-29-22.875716.json,migtissera/Tess-v2.5-Phi-3-medium-128k-14B,False,13960238080,0.932563998838669,0.7178734160253728,0.7793471228083411,0.7161364027810784 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/migtissera/Trinity-2-Codestral-22B/results_2024-10-16T06-41-27.696700.json,migtissera/Trinity-2-Codestral-22B,False,22247319552,0.9113496854193482,0.8099367812429756,0.8212524428832806,0.8412023458132267 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/migtissera/Trinity-2-Codestral-22B-v0.2/results_2024-10-16T05-33-11.892959.json,migtissera/Trinity-2-Codestral-22B-v0.2,False,22247319552,0.9105384841944013,0.7839869368950081,0.7766103692065986,0.8449095547309833 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistral-community/Mistral-7B-v0.2/results_2024-04-17T02-58-10.191549.json,mistral-community/Mistral-7B-v0.2,False,7241732096,0.8769932005108924,0.6693608751529616,0.47700641417918904,0.7787384922608687 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistral-community/Mixtral-8x22B-Instruct-v0.1-4bit/results_2024-04-23T12-47-11.611190.json,mistral-community/Mixtral-8x22B-Instruct-v0.1-4bit,False,140630071296,0.9399136714864611,0.8303945176300034,0.8114423851732474,0.8381920828656453 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistral-community/Mixtral-8x22B-v0.1-4bit/results_2024-04-15T05-18-29.525677.json,mistral-community/Mixtral-8x22B-v0.1-4bit,False,140620634112,0.9190690293573129,0.8288308244321996,0.8503586787361985,0.8138917557606677 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Codestral-22B-v0.1/results_2024-07-03T01-40-04.986395.json,mistralai/Codestral-22B-v0.1,False,22247282688,0.9125769727553994,0.7932190272442804,0.8327673150149223,0.8253547481677842 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Ministral-8B-Instruct-2410/results_2024-12-03T16-09-00.165772.json,mistralai/Ministral-8B-Instruct-2410,False,8019808256,0.9235725320171582,0.7461224853968654,0.7690598312683299,0.8802302570746368 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Mistral-7B-Instruct-v0.1/results_2024-05-29T09-26-28.335343.json,mistralai/Mistral-7B-Instruct-v0.1,False,7241732096,0.8680148326073658,0.658129383494101,0.604143159962043,0.7651799532417927 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Mistral-7B-Instruct-v0.2/results_2024-02-22T04-38-02.592138.json,mistralai/Mistral-7B-Instruct-v0.2,False,7241732096,0.9051791381281589,0.743171027388709,0.6610295761007225,0.8121063440638714 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Mistral-7B-Instruct-v0.3/results_2024-05-23T00-44-16.674482.json,mistralai/Mistral-7B-Instruct-v0.3,False,7248023552,0.9027027027027027,0.7942968956652824,0.6866859299725274,0.8343452944711321 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Mistral-7B-v0.1/results_2024-02-06T13-10-21.397647.json,mistralai/Mistral-7B-v0.1,False,7241732096,0.5483129429185911,0.578493410475696,0.43917169974115616,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Mistral-7B-v0.1/results_2024-02-21T18-21-24.437019.json,mistralai/Mistral-7B-v0.1,False,7241732096,0.8891760520923071,0.6199587479623665,0.48,0.7673143331406637 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Mistral-7B-v0.3/results_2024-05-22T23-44-47.748153.json,mistralai/Mistral-7B-v0.3,False,7248023552,0.8790493645404561,0.6748507212873428,0.4777509957775612,0.778367362714571 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Mistral-Nemo-Base-2407/results_2024-07-24T04-01-50.428826.json,mistralai/Mistral-Nemo-Base-2407,False,12247782400,0.9115228100235597,0.7673981003009931,0.6208649279226731,0.7780596068484464 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Mistral-Nemo-Instruct-2407/results_2024-07-24T02-43-26.936605.json,mistralai/Mistral-Nemo-Instruct-2407,False,12247782400,0.9321884109145206,0.7517568673754127,0.8234480202211933,0.779598729750832 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Mistral-Small-24B-Base-2501/results_2025-01-31T18-21-09.888586.json,mistralai/Mistral-Small-24B-Base-2501,False,23572403200,0.9088638162572613,0.8149746517644091,0.6470346507771108,0.8216240151724024 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Mistral-Small-24B-Instruct-2501/results_2025-01-31T20-05-31.526259.json,mistralai/Mistral-Small-24B-Instruct-2501,False,23572403200,0.937450306390692,0.7610716393662056,0.7962054061973483,0.8736445802550317 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Mixtral-8x22B-Instruct-v0.1/results_2024-05-24T09-58-30.383236.json,mistralai/Mixtral-8x22B-Instruct-v0.1,False,140630071296,0.9407529364400987,0.8416653538742355,0.8073234524847428,0.8495726211789084 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Mixtral-8x7B-Instruct-v0.1/results_2024-02-18T18-26-33.687924.json,mistralai/Mixtral-8x7B-Instruct-v0.1,False,46702792704,0.9264254873898037,0.8273826913007483,0.8009434691825845,0.7848550686238143 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Mixtral-8x7B-v0.1/results_2024-02-10T03-25-53.024286.json,mistralai/Mixtral-8x7B-v0.1,False,46702792704,0.6646069030555132,0.49872071984375066,0.5026086956521739,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mistralai/Mixtral-8x7B-v0.1/results_2024-02-18T23-03-35.536475.json,mistralai/Mixtral-8x7B-v0.1,False,46702792704,0.9103413524823818,0.7870598909460976,0.4830369649343558,0.6929673616240137 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mlabonne/AlphaMonarch-7B/results_2024-05-19T05-10-09.581920.json,mlabonne/AlphaMonarch-7B,False,7241732096,0.5770894729747643,0.566431547018829,0.5282390058211012,0.6660653746489624 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mlabonne/Beyonder-4x7B-v3/results_2024-05-26T14-45-26.008679.json,mlabonne/Beyonder-4x7B-v3,False,24153690112,0.6255595602811317,0.5739332533595589,0.5516489252623926,0.6746769273926633 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mlabonne/ChimeraLlama-3-8B-v3/results_2024-08-09T01-36-08.043461.json,mlabonne/ChimeraLlama-3-8B-v3,False,8030261248,0.9170391560879672,0.7767578927661603,0.743121195311282,0.8279644376879208 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mlabonne/Daredevil-8B-abliterated/results_2024-05-27T13-40-17.084359.json,mlabonne/Daredevil-8B-abliterated,False,8030261248,0.926037098632726,0.7804713793394579,0.7579701787626658,0.8422453298022874 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mlabonne/FineLlama-3.1-8B/results_2024-09-01T04-02-48.724980.json,mlabonne/FineLlama-3.1-8B,False,8030261248,0.9046579667406769,0.786753946311094,0.4396551724137931,0.766040657243156 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mlabonne/Llama-3-8B-Instruct-abliterated-dpomix/results_2024-05-27T14-44-42.467656.json,mlabonne/Llama-3-8B-Instruct-abliterated-dpomix,False,8030261248,0.9136602188080808,0.7676336804192703,0.7143293906330614,0.8645687620632841 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated/results_2024-09-11T01-44-59.225606.json,mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated,False,8030261248,0.9214761710010344,0.7273403262027243,0.7545269436365831,0.8634654749936048 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mlabonne/Monarch-7B/results_2024-02-26T20-44-34.635482.json,mlabonne/Monarch-7B,False,7241732096,0.9240171227138955,0.7697959689241017,0.7656618319975595,0.8376055098130599 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mlabonne/NeuralDaredevil-8B-abliterated/results_2024-05-29T12-03-13.060846.json,mlabonne/NeuralDaredevil-8B-abliterated,False,8030261248,0.9272735526778723,0.7793785227137924,0.7340141621890901,0.8456073425047335 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mlabonne/NeuralMonarch-7B/results_2024-05-21T02-25-10.601922.json,mlabonne/NeuralMonarch-7B,False,7241732096,0.5823519371906469,0.5665781323229131,0.5272169690774342,0.6688152504087473 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/monilouise/opt125M_portuguese/results_2024-02-07T23-58-49.602518.json,monilouise/opt125M_portuguese,False,125239296,0.3333333333333333,0.02570115105018438,0.28426999702646444,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/monilouise/opt125M_portuguese/results_2024-03-08T04-26-59.061295.json,monilouise/opt125M_portuguese,False,125239296,0.2442578766685356,0.015992331599555404,0.27878787878787875,0.2299778028390885 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/monsterapi/Llama-3_1-8B-Instruct-orca-ORPO/results_2025-02-04T16-26-20.906343.json,monsterapi/Llama-3_1-8B-Instruct-orca-ORPO,False,8057524224,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mosaicml/mpt-30b/results_2024-02-09T05-24-45.198006.json,mosaicml/mpt-30b,False,29957200896,0.5531464415296898,0.3439139332088446,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mosaicml/mpt-30b/results_2024-08-12T05-11-21.745805.json,mosaicml/mpt-30b,False,29957200896,0.6040744959014421,0.5316424628897842,0.4396551724137931,0.838076159721676 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mosaicml/mpt-7b/results_2024-02-09T05-01-56.623731.json,mosaicml/mpt-7b,False,6649286656,0.34145160422532683,0.18509682219401635,0.4605164949992536,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mosaicml/mpt-7b/results_2024-04-16T15-25-08.488964.json,mosaicml/mpt-7b,False,6649286656,0.7649351826517844,0.2279546636099929,0.446593203434771,0.7886907843528153 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mosaicml/mpt-7b-8k/results_2024-02-09T06-08-56.199589.json,mosaicml/mpt-7b-8k,False,6649286656,0.46560106560106557,0.17918361515441739,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/mosaicml/mpt-7b-8k/results_2024-04-16T17-02-27.175833.json,mosaicml/mpt-7b-8k,False,6649286656,0.8111566261572349,0.2528414822324891,0.4396551724137931,0.7482646670430639 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MTSAIR/multi_verse_model/results_2024-05-26T23-29-38.848309.json,MTSAIR/multi_verse_model,False,7241732096,0.9225619834710744,0.7909251769543553,0.794365153254454,0.8432720746227155 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MulaBR/Mula-4x160-v0.1/results_2024-04-22T00-05-24.255163.json,MulaBR/Mula-4x160-v0.1,False,417258240,0.335683441456502,0.11349165436666529,0.4396551724137931,0.41502718891863716 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/MulaBR/Mula-8x160-v0.1/results_2024-05-08T21-44-31.351826.json,MulaBR/Mula-8x160-v0.1,False,747596544,0.33568542924880684,0.04732817202670604,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/natong19/Mistral-Nemo-Instruct-2407-abliterated/results_2024-10-02T04-37-08.034005.json,natong19/Mistral-Nemo-Instruct-2407-abliterated,False,12247782400,0.9285000788602071,0.7298741419125692,0.8246637558968182,0.7561303732384439 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/natong19/Mistral-Nemo-Instruct-2407-abliterated/results_2024-10-02T04-38-27.639018.json,natong19/Mistral-Nemo-Instruct-2407-abliterated,False,12247782400,0.9285000788602071,0.7298741419125692,0.8246637558968182,0.7561303732384439 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/natong19/Qwen2-7B-Instruct-abliterated/results_2024-06-19T14-18-57.119242.json,natong19/Qwen2-7B-Instruct-abliterated,False,7615616512,0.9362745098039216,0.7586495215882257,0.750904832872046,0.7756938741296732 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nbeerbower/gemma2-gutenberg-27B/results_2024-09-19T03-43-58.258903.json,nbeerbower/gemma2-gutenberg-27B,False,27227128320,0.8263497179693795,0.4506424390957356,0.5992314436547466,0.8146115971429779 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nbeerbower/gemma2-gutenberg-27B/results_2024-09-19T03-47-20.971043.json,nbeerbower/gemma2-gutenberg-27B,False,27227128320,0.8263497179693795,0.4506424390957356,0.5992314436547466,0.8146115971429779 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nbeerbower/mistral-nemo-wissenschaft-12B/results_2025-02-05T00-28-59.943310.json,nbeerbower/mistral-nemo-wissenschaft-12B,False,12247782400,0.9268789751522479,0.726265639709004,0.8253670727897532,0.7556200952674685 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Nexusflow/Starling-LM-7B-beta/results_2024-04-01T21-15-54.379246.json,Nexusflow/Starling-LM-7B-beta,False,7241748480,0.9256528007689433,0.8246749931266709,0.7748688218404758,0.8311091883257347 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicholasKluge/Aira-2-portuguese-124M/results_2024-02-06T18-59-18.399401.json,nicholasKluge/Aira-2-portuguese-124M,False,124442880,0.3333333333333333,0.023587414997509645,0.12578616352201258,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicholasKluge/Aira-2-portuguese-124M/results_2024-02-18T06-03-33.775333.json,nicholasKluge/Aira-2-portuguese-124M,False,124442880,0.42517266171376283,0.02922061024865183,0.057971014492753624,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicholasKluge/Aira-2-portuguese-1B7/results_2024-02-06T20-19-33.747613.json,nicholasKluge/Aira-2-portuguese-1B7,False,1722007552,0.02915848253477538,0.04201899428767068,0.02354154211397329,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicholasKluge/Aira-2-portuguese-1B7/results_2024-02-18T09-32-23.986529.json,nicholasKluge/Aira-2-portuguese-1B7,False,1722007552,0.02531796054317192,0.010369387902484104,0.019896805202148714,0.16982016982016981 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicholasKluge/Aira-2-portuguese-560M/results_2024-02-06T19-22-44.830237.json,nicholasKluge/Aira-2-portuguese-560M,False,559013888,0.010068892421833599,,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicholasKluge/Aira-2-portuguese-560M/results_2024-02-18T06-49-24.216224.json,nicholasKluge/Aira-2-portuguese-560M,False,559013888,0.0,0.0,0.0,0.011954022988505746 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicholasKluge/baseline/results_2024-05-07T02-05-45.266673.json,nicholasKluge/baseline,False,747596544,0.33514630774633175,0.0,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicholasKluge/TeenyTinyLlama-160m/results_2024-02-25T07-53-21.405861.json,nicholasKluge/TeenyTinyLlama-160m,False,162417408,0.539728635863353,0.002356037361961063,0.4396551724137931,0.3692355648458845 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicholasKluge/TeenyTinyLlama-460m/results_2024-02-25T09-02-16.447511.json,nicholasKluge/TeenyTinyLlama-460m,False,468239360,0.5360816542791849,0.13004763756618948,0.4640538365450605,0.3358561762664816 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicholasKluge/TeenyTinyLlama-460m-Chat/results_2024-02-25T09-45-35.395101.json,nicholasKluge/TeenyTinyLlama-460m-Chat,False,468243456,0.4377495864867047,0.04519453586348976,0.3399609444345819,0.3349186726374015 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicolasdec/Cabra/results_2024-02-06T16-32-19.515237.json,nicolasdec/Cabra,False,6738415616,0.6031549316259776,0.07928346438495378,0.47034320848362593,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicolasdec/Cabra/results_2024-02-25T17-40-20.932980.json,nicolasdec/Cabra,False,6738415616,0.8111893155258765,0.29375427973749235,0.4728966877507991,0.8170891037366084 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicolasdec/cabra13b/results_2024-02-09T18-05-05.112496.json,nicolasdec/cabra13b,False,13015864320,0.7303144625479399,0.22425099777612584,0.5499330669689338,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicolasdec/cabra13b/results_2024-02-17T23-07-55.801702.json,nicolasdec/cabra13b,False,13015864320,0.8554620405358984,0.5719364486772344,0.45449901481427424,0.7777544154751892 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicolasdec/CabraMistral7b/results_2024-04-02T03-35-22.746080.json,nicolasdec/Cabramistral7b,False,7241732096,0.9007017401533282,0.7248226511821165,0.647999490316004,0.8307008241707131 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicolasdec/CabraMistral7b/results_2024-04-02T10-38-23.288377.json,nicolasdec/CabraMistral7b,False,7241732096,0.9027403846153846,0.7224760231184958,0.64349376114082,0.8314517624968287 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicolasdec/CabraMistral7b/results_2024-04-19T17-40-12.612095.json,nicolasdec/Cabramistral7b,False,7241732096,0.9027403846153846,0.7224760231184958,0.64349376114082,0.8314517624968287 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicolasdec/CabraMistral7b-0.4/results_2024-02-25T11-07-27.967744.json,nicolasdec/CabraMistral7b-0.4,False,7241732096,0.695368369677116,0.3959797457650186,0.4396551724137931,0.7845252167060677 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicolasdec/CabraQwen14b/results_2024-04-02T13-21-21.355865.json,nicolasdec/CabraQwen14b,False,14163010560,0.9142128804834304,0.8099427333727796,0.4611447520184544,0.7937678467612272 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicolasdec/CabraQwen14b/results_2024-04-02T14-07-49.242860.json,nicolasdec/CabraQwen14b,False,14163010560,0.9142110480309993,0.8085402540491559,0.4605164949992536,0.7931819181905618 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nicolasdec/CabraQwen7b/results_2024-04-02T03-36-31.835554.json,nicolasdec/CabraQwen7b,False,7718948864,0.8851585879218251,0.7617055030689699,0.5779905859438402,0.7631994433746268 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nisten/Biggie-SmoLlm-0.15B-Base/results_2024-08-11T18-22-45.329968.json,nisten/Biggie-SmoLlm-0.15B-Base,False,152215488,0.3333333333333333,0.020777363974771167,0.4396551724137931,0.34368463405865285 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nisten/Biggie-SmoLlm-0.15B-Base/results_2024-08-11T18-22-45.332134.json,nisten/Biggie-SmoLlm-0.15B-Base,False,152215488,0.3333333333333333,0.020777363974771167,0.4396551724137931,0.34368463405865285 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/nisten/franqwenstein-35b/results_2024-12-04T15-18-56.483770.json,nisten/franqwenstein-35b,False,34714297344,0.9443989816633944,0.83976286129967,0.8620330348873109,0.8922626083530338 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NLPark/AnFeng_v3.1-Avocet/results_2024-08-07T01-36-03.924303.json,NLPark/AnFeng_v3.1-Avocet,False,34392587264,0.910014436186708,0.8080882984014623,0.81625848463485,0.8508098891730606 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NLPark/AnFeng_v3_Avocet/results_2024-05-25T07-09-02.650733.json,NLPark/AnFeng_v3_Avocet,False,34980831232,0.6453682460743775,0.1094044556850774,0.0,0.4521948413403378 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Nos-PT/Carvalho_pt-gl-1.3B/results_2024-06-29T01-31-46.614292.json,Nos-PT/Carvalho_pt-gl-1.3B,False,1315723264,0.36743639380530974,0.13874308801011379,0.4396551724137931,0.39197789682567186 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Nos-PT/Llama-Carvalho-PT/results_2025-04-04T04-18-37.571793.json,Nos-PT/Llama-Carvalho-PT,False,8030261248,0.8749812243524655,0.7569955132231747,0.4396551724137931,0.7692906208084841 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Nos-PT/Llama-Carvalho-PT-GL/results_2025-04-04T00-18-25.576865.json,Nos-PT/Llama-Carvalho-PT-GL,False,8030261248,0.8929720702385808,0.781839864012589,0.6587838419304315,0.8283297331556958 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NotAiLOL/Yi-1.5-dolphin-9B/results_2024-05-20T01-10-12.037268.json,NotAiLOL/Yi-1.5-dolphin-9B,False,8829407232,0.8760640976552467,0.7034632549590849,0.5722163805950391,0.8610610921948665 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NousResearch/Hermes-2-Pro-Llama-3-8B/results_2024-05-07T04-01-33.854422.json,NousResearch/Hermes-2-Pro-Llama-3-8B,False,8030523392,0.9223739628332301,0.7575480918675715,0.7486659964426572,0.821316847945847 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NousResearch/Hermes-2-Theta-Llama-3-8B/results_2024-05-27T07-29-46.242035.json,NousResearch/Hermes-2-Theta-Llama-3-8B,False,8030261248,0.9207473491333994,0.7468943971678241,0.7622538924128841,0.7992468869348173 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NousResearch/Hermes-3-Llama-3.1-8B/results_2024-09-02T01-33-16.513631.json,NousResearch/Hermes-3-Llama-3.1-8B,False,8030261248,0.9132298061875527,0.7253208984405229,0.7135850719390003,0.8535677677656227 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NousResearch/Meta-Llama-3-8B/results_2024-05-21T19-05-24.569014.json,NousResearch/Meta-Llama-3-8B,False,8030261248,0.8951198939631281,0.766493362127982,0.4396551724137931,0.7562902556274187 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NousResearch/Nous-Capybara-34B/results_2024-05-16T07-59-45.028987.json,NousResearch/Nous-Capybara-34B,False,34388917248,0.9007100934823724,0.757100596654299,0.7731239092495636,0.7408626005155765 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NousResearch/Nous-Hermes-13b/results_2024-02-28T05-28-55.109832.json,NousResearch/Nous-Hermes-13b,False,13015874560,0.6630418610067084,0.5166061482295778,0.6245167597351622,0.7448674633758359 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NousResearch/Nous-Hermes-2-Mistral-7B-DPO/results_2024-02-27T02-51-13.508742.json,NousResearch/Nous-Hermes-2-Mistral-7B-DPO,False,7241748480,0.9021970801589174,0.6915650379510005,0.7138364779874213,0.7767581619154933 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO/results_2024-03-06T22-07-16.186340.json,NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO,False,46702809088,0.9011405575094769,0.7346929104749711,0.7626485982066783,0.7640680874353314 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO/results_2024-03-07T12-23-27.230508.json,NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO,False,46702809088,0.9023645725471283,0.734857329244095,0.7498307874026198,0.7666031472700376 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NousResearch/Nous-Hermes-2-SOLAR-10.7B/results_2024-02-27T18-41-43.250837.json,NousResearch/Nous-Hermes-2-SOLAR-10.7B,False,10731540480,0.9177576504248857,0.797009717894243,0.8008578431372549,0.8507132845086443 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NousResearch/Nous-Hermes-2-Yi-34B/results_2024-02-27T03-50-28.297314.json,NousResearch/Nous-Hermes-2-Yi-34B,False,34388917248,0.9215044447012628,0.7985401560561216,0.7605236777394121,0.7703803469511286 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NOVA-vision-language/GlorIA-1.3B/results_2024-03-07T22-34-35.217921.json,NOVA-vision-language/GlorIA-1.3B,False,1315577856,0.0,0.023212602251989234,0.00390625,0.004265433443903698 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NTQAI/Nxcode-CQ-7B-orpo/results_2024-08-11T21-22-31.062573.json,NTQAI/Nxcode-CQ-7B-orpo,False,7250284544,0.8668091431605842,0.6966418242642785,0.4471267110923455,0.7517084027692342 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NucleusAI/nucleus-22B-token-500B/results_2024-02-14T13-14-20.585602.json,NucleusAI/nucleus-22B-token-500B,False,21827959296,0.33965642015109887,0.14105865232500697,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/NucleusAI/nucleus-22B-token-500B/results_2024-04-03T18-19-57.556537.json,NucleusAI/nucleus-22B-token-500B,False,21827959296,0.3333333333333333,0.18130129349722302,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Oburaco/juridic-falcon-7b/results_2024-05-27T21-44-42.226357.json,Oburaco/juridic-falcon-7b,False,6924080000,0.5892908131346348,0.019937421587094828,0.4396551724137931,0.7484848438318135 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Oburaco/llama2-qlora-finetunined-ptbr/results_2024-05-27T20-27-01.506989.json,Oburaco/llama2-qlora-finetunined-ptbr,False,6771970048,0.6353154692003844,0.3036032916828423,0.5303967191517392,0.7372127835097078 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Oburaco/ptbr-falcon-7b/results_2024-05-27T17-07-39.276986.json,Oburaco/ptbr-falcon-7b,False,6926439296,0.5193803871944144,0.0830274948506502,0.4396551724137931,0.5267321985135459 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Oburaco/ptbr-falcon-7b-v3/results_2024-05-27T18-44-47.334916.json,Oburaco/ptbr-falcon-7b-v3,False,6926439296,0.5442507626719338,0.05084063725467101,0.4396551724137931,0.47451878248737356 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OliveiraJLT/Sagui-7B-Instruct-v0.1/results_2024-07-18T10-05-57.915014.json,OliveiraJLT/Sagui-7B-Instruct-v0.1,False,6738415616,0.7116215451710273,0.03155319076723917,0.5805353659391029,0.4646385620718362 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OliveiraJLT/Sagui-7B-Instruct-v0.1/results_2024-07-18T10-05-57.992253.json,OliveiraJLT/Sagui-7B-Instruct-v0.1,False,6738415616,0.7116215451710273,0.03155319076723917,0.5805353659391029,0.4646385620718362 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OliveiraJLT/Sagui-7B-Instruct-v0.1/results_2024-07-18T10-06-00.459222.json,OliveiraJLT/Sagui-7B-Instruct-v0.1,False,6738415616,0.7116215451710273,0.03155319076723917,0.5805353659391029,0.4646385620718362 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openai-community/gpt2-large/results_2024-02-09T01-12-14.059849.json,openai-community/gpt2-large,False,774030080,0.3333333333333333,,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openai-community/gpt2-large/results_2024-04-16T13-23-43.768045.json,openai-community/gpt2-large,False,774030080,0.3333333333333333,0.16216551599232715,0.4396551724137931,0.33459895328341466 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openai-community/gpt2-medium/results_2024-02-09T00-38-03.057064.json,openai-community/gpt2-medium,False,354823168,0.3333333333333333,0.008744289928817432,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openai-community/gpt2-medium/results_2024-04-16T12-06-01.268166.json,openai-community/gpt2-medium,False,354823168,0.33801352008051505,0.16260640310673247,0.4396551724137931,0.346461375361181 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openai-community/gpt2-xl/results_2024-02-09T01-34-08.460486.json,openai-community/gpt2-xl,False,1557611200,0.3333333333333333,0.039738468599691254,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openai-community/gpt2-xl/results_2024-04-16T13-54-03.146810.json,openai-community/gpt2-xl,False,1557611200,0.3442647527312932,0.010972719883957299,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openai-community/openai-gpt/results_2024-02-08T23-49-26.472276.json,openai-community/openai-gpt,False,116534784,0.0,0.0797566804202465,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openai-community/openai-gpt/results_2024-04-16T11-44-49.742462.json,openai-community/openai-gpt,False,116534784,0.003257328990228013,0.042727458076029255,0.0,0.0014265335235378032 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OpenBuddy/openbuddy-llama3.1-8b-v22.3-131k/results_2024-10-04T04-17-28.382369.json,OpenBuddy/openbuddy-llama3.1-8b-v22.3-131k,False,8030261248,0.8525421062839115,0.7740544972249987,0.7150974025974026,0.8735636230604032 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OpenBuddy/openbuddy-qwen1.5-32b-v21.1-32k/results_2024-06-12T17-26-31.366416.json,OpenBuddy/openbuddy-qwen1.5-32b-v21.1-32k,False,32512218112,0.9133644488776482,0.7737023206349624,0.7804272230501739,0.8941039621063908 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OpenBuddy/openbuddy-qwen1.5-32b-v21.2-32k/results_2024-06-13T03-15-43.142431.json,OpenBuddy/openbuddy-qwen1.5-32b-v21.2-32k,False,32512218112,0.9221381866393468,0.8036205393734592,0.8168992136555427,0.8860083727003879 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OpenBuddy/openbuddy-qwen2.5llamaify-14b-v23.1-200k/results_2024-10-04T04-07-32.668972.json,OpenBuddy/openbuddy-qwen2.5llamaify-14b-v23.1-200k,False,14769689600,0.87251507886164,0.7813303083235066,0.6292421694999688,0.8532431415833708 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OpenBuddy/openbuddy-qwen2.5llamaify-7b-v23.1-200k/results_2024-10-13T01-35-20.349000.json,OpenBuddy/openbuddy-qwen2.5llamaify-7b-v23.1-200k,False,7615487488,0.9178668153377105,0.7803116465290667,0.791095879513395,0.7848941181688438 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OpenBuddy/openbuddy-qwq-32b-v24.1-200k/results_2025-01-31T15-43-40.673905.json,OpenBuddy/openbuddy-qwq-32b-v24.1-200k,False,32763876352,0.9394606560434073,0.8462136153703603,0.8264160673339902,0.9149270369127809 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openchat/openchat-3.5-0106/results_2024-02-08T20-30-27.905193.json,openchat/openchat-3.5-0106,False,7241748480,0.9211600123149934,0.7374716904227555,0.7951201656662747,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openchat/openchat-3.5-0106/results_2024-02-21T21-51-39.369217.json,openchat/openchat-3.5-0106,False,7241748480,0.9272805921368836,0.8316751955429375,0.7920758364261121,0.7930509518656714 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openchat/openchat-3.5-0106-gemma/results_2024-05-20T02-40-24.454178.json,openchat/openchat-3.5-0106-gemma,False,8537680896,0.9399059653001771,0.7910602658449247,0.8425717427733557,0.9032788470838573 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openchat/openchat-3.6-8b-20240522/results_2024-05-25T08-56-51.584208.json,openchat/openchat-3.6-8b-20240522,False,8030261248,0.9262993453513928,0.7794747802276402,0.7363888888888889,0.8551991680384323 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openchat/openchat-3.6-8b-20240522/results_2024-05-26T09-38-49.955889.json,openchat/openchat-3.6-8b-20240522,False,8030261248,0.9271239460213057,0.7807795684477881,0.7479884024094219,0.8543802303140403 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OpenLLM-France/Claire-7B-0.1/results_2024-02-08T19-59-33.207703.json,OpenLLM-France/Claire-7B-0.1,False,6921720704,0.3333333333333333,0.043451880521346076,0.4471267110923455,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OpenLLM-France/Claire-7B-0.1/results_2024-04-16T02-18-49.342048.json,OpenLLM-France/Claire-7B-0.1,False,6921720704,0.4166035250383718,0.10146422112236213,0.44552575932333716,0.3349186726374015 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OpenLLM-France/Claire-Mistral-7B-0.1/results_2024-02-08T18-48-18.223603.json,OpenLLM-France/Claire-Mistral-7B-0.1,False,7241732096,0.3930811717227137,0.5428546315437224,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OpenLLM-France/Claire-Mistral-7B-0.1/results_2024-04-16T01-27-34.625162.json,OpenLLM-France/Claire-Mistral-7B-0.1,False,7241732096,0.8558405766376935,0.6134554484624953,0.4396551724137931,0.7566238658528611 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openlm-research/open_llama_13b/results_2024-02-16T11-22-43.611615.json,openlm-research/open_llama_13b,False,13015864320,0.5467070837305316,0.12834096349450685,0.4611447520184544,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openlm-research/open_llama_13b/results_2024-02-17T13-43-33.351925.json,openlm-research/open_llama_13b,False,13015864320,0.50833308023682,0.21872204239614385,0.4471267110923455,0.4646385620718362 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openlm-research/open_llama_3b/results_2024-02-16T07-23-49.395940.json,openlm-research/open_llama_3b,False,3426473600,0.3333333333333333,0.014951178957756552,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openlm-research/open_llama_3b/results_2024-02-17T09-33-21.785739.json,openlm-research/open_llama_3b,False,3426473600,0.35297899277093914,0.05530368768431126,0.4396551724137931,0.38836470532454603 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openlm-research/open_llama_3b_v2/results_2024-02-16T09-14-10.452122.json,openlm-research/open_llama_3b_v2,False,3426473600,0.45744669644470837,0.008272566397281897,0.43771626297577854,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openlm-research/open_llama_3b_v2/results_2024-02-17T10-33-05.152124.json,openlm-research/open_llama_3b_v2,False,3426473600,0.33424046164366045,0.06301379230313081,0.4386873920552677,0.33459895328341466 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openlm-research/open_llama_7b/results_2024-02-16T10-24-26.090061.json,openlm-research/open_llama_7b,False,6738415616,0.3333333333333333,0.17517758740700018,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openlm-research/open_llama_7b/results_2024-02-17T11-32-23.927613.json,openlm-research/open_llama_7b,False,6738415616,0.35551526385531756,0.16402785934407932,0.4396551724137931,0.3349186726374015 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openlm-research/open_llama_7b_v2/results_2024-02-16T11-06-23.197438.json,openlm-research/open_llama_7b_v2,False,6738415616,0.4061672924232166,0.00998241315939068,0.4971032623376896,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/openlm-research/open_llama_7b_v2/results_2024-02-17T12-39-02.708744.json,openlm-research/open_llama_7b_v2,False,6738415616,0.6480758670160722,0.03824405594408843,0.5221900754232328,0.7441047463175122 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Orion-zhen/Qwen2.5-7B-Instruct-Uncensored/results_2024-10-20T01-45-09.693782.json,Orion-zhen/Qwen2.5-7B-Instruct-Uncensored,False,7615616512,0.9374994889577324,0.7944174834061616,0.7778487466523234,0.7859209826965098 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Orion-zhen/Qwen2.5-7B-Instruct-Uncensored/results_2024-10-20T01-45-09.738496.json,Orion-zhen/Qwen2.5-7B-Instruct-Uncensored,False,7615616512,0.9374994889577324,0.7944174834061616,0.7778487466523234,0.7859209826965098 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OrionStarAI/Orion-14B-Base/results_2024-02-13T19-40-35.556453.json,OrionStarAI/Orion-14B-Base,False,14498703360,0.6307200282684031,0.5853055022861218,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OrionStarAI/Orion-14B-Base/results_2024-04-03T01-15-16.610234.json,OrionStarAI/Orion-14B-Base,False,14498703360,0.9030903950791537,0.7713114437383479,0.4396551724137931,0.5878495043212718 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/OrionStarAI/OrionStar-Yi-34B-Chat/results_2024-06-15T03-09-25.428998.json,OrionStarAI/OrionStar-Yi-34B-Chat,False,34388917248,0.9170600914778403,0.796949606031967,0.7839813891658358,0.8443962929338467 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/paulml/OGNO-7B/results_2024-02-26T21-46-38.735296.json,paulml/OGNO-7B,False,7241732096,0.9195202213253991,0.777306832779747,0.7852659398744631,0.8146674347869539 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/pedrogengo/gemma-ptbr/results_2024-05-19T06-08-26.193367.json,pedrogengo/gemma-ptbr,False,2515978240,0.3333333333333333,0.0333526073052867,0.5199346955699237,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/pfnet/plamo-13b/results_2024-02-08T14-27-00.255480.json,pfnet/plamo-13b,False,13099545600,0.5117413644961335,0.0340738010760964,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/pfnet/plamo-13b/results_2024-05-22T06-27-47.187716.json,pfnet/plamo-13b,False,13099545600,0.5933143956158916,0.13875118119447466,0.43917169974115616,0.4059003157657368 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/pierreguillou/gpt2-small-portuguese/results_2024-02-07T21-11-02.837811.json,pierreguillou/gpt2-small-portuguese,False,124439808,0.3333333333333333,0.023733232880953473,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/pierreguillou/gpt2-small-portuguese/results_2024-03-08T03-48-50.928791.json,pierreguillou/gpt2-small-portuguese,False,124439808,0.33586720612107884,0.03435992511506735,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/PORTULAN/gervasio-7b-portuguese-ptbr-decoder/results_2024-03-07T22-39-33.153068.json,PORTULAN/gervasio-7b-portuguese-ptbr-decoder,False,6738415616,0.8315268026535532,0.695518625458929,0.18589914951337116,0.5380422946087232 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/PORTULAN/gervasio-7b-portuguese-ptpt-decoder/results_2024-03-08T02-58-56.846301.json,PORTULAN/gervasio-7b-portuguese-ptpt-decoder,False,6738415616,0.85320443811568,0.7909512765742471,0.6221487631987646,0.617427881406052 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/premai-io/prem-1B-chat/results_2024-07-24T07-14-48.642589.json,premai-io/prem-1B-chat,False,1100064768,0.3537827992186262,0.0015225595375366486,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/gemma-2-9b-it-DPO/results_2024-10-07T05-41-20.269484.json,princeton-nlp/gemma-2-9b-it-DPO,False,9241705984,0.9337216018482866,0.7709822151499908,0.6694711538461537,0.8851515282024779 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/gemma-2-9b-it-DPO/results_2025-05-07T16-33-20.423477.json,princeton-nlp/gemma-2-9b-it-DPO,False,0,0.93453985367732,0.7694771494118819,0.6734320952067204,0.8851044011169493 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/gemma-2-9b-it-SimPO/results_2024-08-11T19-34-03.486958.json,princeton-nlp/gemma-2-9b-it-SimPO,False,9241705984,0.9382389870905168,0.7781569502589806,0.7045454545454546,0.8976498838719835 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-8B-ProLong-512k-Instruct/results_2024-09-21T05-00-53.028570.json,princeton-nlp/Llama-3-8B-ProLong-512k-Instruct,False,8030261248,0.9211600123149934,0.749376027031628,0.7101770443131423,0.8370788715394708 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-8B-ProLong-64k-Base/results_2024-09-21T03-56-02.995738.json,princeton-nlp/Llama-3-8B-ProLong-64k-Base,False,8030261248,0.9244235518560937,0.7692389686380259,0.7225871779207846,0.842471218628472 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-8B-ProLong-64k-Instruct/results_2024-09-21T04-29-42.154658.json,princeton-nlp/Llama-3-8B-ProLong-64k-Instruct,False,8030261248,0.9207236418194675,0.7510972158248586,0.7011494252873564,0.8290775286403318 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Base-8B-SFT/results_2024-10-07T08-15-14.872557.json,princeton-nlp/Llama-3-Base-8B-SFT,False,8030261248,0.8770865683229814,0.7422721589146833,0.2651740173472673,0.84472049689441 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Base-8B-SFT-CPO/results_2024-10-17T07-35-19.922338.json,princeton-nlp/Llama-3-Base-8B-SFT-CPO,False,8030261248,0.9037275775607339,0.7650704792433256,0.4137781659884815,0.8483622303422681 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Base-8B-SFT-DPO/results_2024-12-07T04-00-43.491525.json,princeton-nlp/Llama-3-Base-8B-SFT-DPO,False,8030261248,0.9178150146340144,0.7097666460699621,0.5334528807814274,0.8265180569816712 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Base-8B-SFT-IPO/results_2024-10-11T01-39-18.193786.json,princeton-nlp/Llama-3-Base-8B-SFT-IPO,False,8030261248,0.9162188986618296,0.7038399659091228,0.5827750402093081,0.7968535707431695 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Base-8B-SFT-KTO/results_2024-10-04T01-46-44.332879.json,princeton-nlp/Llama-3-Base-8B-SFT-KTO,False,8030261248,0.9150099881424192,0.7107889165690748,0.612807148175726,0.8327993344919324 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Base-8B-SFT-ORPO/results_2024-10-04T02-57-15.821892.json,princeton-nlp/Llama-3-Base-8B-SFT-ORPO,False,8030261248,0.9007196790939123,0.7073899885874841,0.3946266449590138,0.8377875404701711 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Base-8B-SFT-RDPO/results_2024-10-04T01-41-17.587145.json,princeton-nlp/Llama-3-Base-8B-SFT-RDPO,False,8030261248,0.9158173076923077,0.7101526567318941,0.5413882367106134,0.8219155154433444 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Base-8B-SFT-RRHF/results_2024-10-11T02-44-35.299155.json,princeton-nlp/Llama-3-Base-8B-SFT-RRHF,False,8030261248,0.8462286496588292,0.6853637543430299,0.2133706130284188,0.7987594010206822 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Base-8B-SFT-SimPO/results_2024-10-02T01-43-07.309826.json,princeton-nlp/Llama-3-Base-8B-SFT-SimPO,False,8030261248,0.9124526142622129,0.6993043161001273,0.5324010128913443,0.8175895765472313 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Base-8B-SFT-SLiC-HF/results_2024-10-11T03-55-07.335985.json,princeton-nlp/Llama-3-Base-8B-SFT-SLiC-HF,False,8030261248,0.9028359120060894,0.7223477802251257,0.4276868893433311,0.8299639583037598 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-CPO/results_2024-10-03T06-15-18.706069.json,princeton-nlp/Llama-3-Instruct-8B-CPO,False,8030261248,0.91241748681093,0.7413359332483361,0.7725788393356969,0.8675353754707225 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-CPO-v0.2/results_2024-10-03T06-02-25.295591.json,princeton-nlp/Llama-3-Instruct-8B-CPO-v0.2,False,8030261248,0.916161792504446,0.7458685139222516,0.694750024290449,0.8711063372717508 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-DPO/results_2024-10-03T07-03-27.273425.json,princeton-nlp/Llama-3-Instruct-8B-DPO,False,8030261248,0.9075953011978164,0.7044183777318505,0.7337967025894853,0.8539087129933345 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-DPO-v0.2/results_2024-10-03T04-28-28.939348.json,princeton-nlp/Llama-3-Instruct-8B-DPO-v0.2,False,8030261248,0.906511737287369,0.7295475285092475,0.5687342056220668,0.8653581718177057 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-IPO/results_2024-10-11T04-59-11.144171.json,princeton-nlp/Llama-3-Instruct-8B-IPO,False,8030261248,0.9195035639155449,0.7253017797778314,0.6885039314176178,0.8756571895258026 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-IPO-v0.2/results_2024-10-11T06-04-33.077855.json,princeton-nlp/Llama-3-Instruct-8B-IPO-v0.2,False,8030261248,0.9002822740638262,0.7240071735716995,0.5790155440414508,0.872142270041036 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-KTO/results_2024-10-03T07-13-34.121902.json,princeton-nlp/Llama-3-Instruct-8B-KTO,False,8030261248,0.9100932121844265,0.701569543919038,0.6734209461422934,0.87060330824549 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-KTO-v0.2/results_2024-10-03T04-28-28.832084.json,princeton-nlp/Llama-3-Instruct-8B-KTO-v0.2,False,8030261248,0.9090651103019919,0.7391640853553003,0.6393631208706186,0.8727532679738562 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-ORPO/results_2024-10-03T08-33-21.561224.json,princeton-nlp/Llama-3-Instruct-8B-ORPO,False,8030261248,0.9116556741942102,0.7170078205237258,0.7491723703718748,0.8734406721229842 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-ORPO/results_2024-10-03T08-35-28.839719.json,princeton-nlp/Llama-3-Instruct-8B-ORPO,False,8030261248,0.9116556741942102,0.7170078205237258,0.7491723703718748,0.8734406721229842 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-ORPO-v0.2/results_2024-10-03T03-24-42.486797.json,princeton-nlp/Llama-3-Instruct-8B-ORPO-v0.2,False,8030261248,0.9161264821473787,0.7459941150404783,0.6554880177052012,0.8673910494590424 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-ORPO-v0.2/results_2024-10-03T03-27-41.208709.json,princeton-nlp/Llama-3-Instruct-8B-ORPO-v0.2,False,8030261248,0.9161264821473787,0.7459941150404783,0.6554880177052012,0.8673910494590424 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-RDPO/results_2024-10-17T03-24-48.994679.json,princeton-nlp/Llama-3-Instruct-8B-RDPO,False,8030261248,0.9043195441802541,0.7028242389204467,0.7242080790467887,0.8432241940469574 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-RDPO-v0.2/results_2024-10-18T03-48-53.926167.json,princeton-nlp/Llama-3-Instruct-8B-RDPO-v0.2,False,8030261248,0.904842328757208,0.7293853336081869,0.5556363636363636,0.8667501780757989 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-RRHF/results_2024-10-08T01-39-08.030156.json,princeton-nlp/Llama-3-Instruct-8B-RRHF,False,8030261248,0.9103413524823818,0.7280440849432734,0.6948528721386589,0.8535384615384616 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-RRHF-v0.2/results_2024-10-08T02-37-24.910845.json,princeton-nlp/Llama-3-Instruct-8B-RRHF-v0.2,False,8030261248,0.924395290030906,0.6810970049712828,0.6778051409553334,0.8771268083994643 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-SimPO/results_2024-05-30T02-34-14.763780.json,princeton-nlp/Llama-3-Instruct-8B-SimPO,False,8030261248,0.9030211936375125,0.7249953009677698,0.5990940435384879,0.831110379359638 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-SimPO/results_2024-10-19T01-44-51.232486.json,princeton-nlp/Llama-3-Instruct-8B-SimPO,False,8030261248,0.9055338303812654,0.7112430937024852,0.7452054153356502,0.8588567395906845 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-SimPO/results_2024-10-19T01-44-51.720817.json,princeton-nlp/Llama-3-Instruct-8B-SimPO,False,8030261248,0.9055338303812654,0.7112430937024852,0.7452054153356502,0.8588567395906845 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-SimPO-v0.2/results_2024-08-11T19-34-03.542478.json,princeton-nlp/Llama-3-Instruct-8B-SimPO-v0.2,False,8030261248,0.896457973466223,0.6991081985044032,0.664955263315919,0.8632511993787806 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-SimPO-v0.2/results_2025-05-07T15-51-57.421344.json,princeton-nlp/Llama-3-Instruct-8B-SimPO-v0.2,False,0,0.8739668044845157,0.6849205578626394,0.6179372835523507,0.8419469928644241 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-SLiC-HF/results_2024-10-08T03-39-14.780417.json,princeton-nlp/Llama-3-Instruct-8B-SLiC-HF,False,8030261248,0.9079252335511221,0.7138919241558832,0.7476738213829266,0.8719153762370823 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Llama-3-Instruct-8B-SLiC-HF-v0.2/results_2024-10-08T04-39-20.490703.json,princeton-nlp/Llama-3-Instruct-8B-SLiC-HF-v0.2,False,8030261248,0.924395290030906,0.6810970049712828,0.6778051409553334,0.8771268083994643 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Base-SFT-CPO/results_2024-10-08T05-41-09.922450.json,princeton-nlp/Mistral-7B-Base-SFT-CPO,False,7241732096,0.8695735610346771,0.6711665077554146,0.733388973895303,0.822709639159549 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Base-SFT-DPO/results_2024-10-08T06-48-28.188515.json,princeton-nlp/Mistral-7B-Base-SFT-DPO,False,7241732096,0.8941989050306216,0.6982344473163888,0.7058374499732827,0.7827980857589536 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Base-SFT-IPO/results_2024-10-08T08-22-24.575645.json,princeton-nlp/Mistral-7B-Base-SFT-IPO,False,7241732096,0.8888461538461538,0.6788057911101084,0.7036832809672127,0.7548972796781197 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Base-SFT-KTO/results_2024-10-18T03-49-07.413834.json,princeton-nlp/Mistral-7B-Base-SFT-KTO,False,7241732096,0.9039947932094774,0.7152295055766035,0.6614751674304303,0.7489550051359846 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Base-SFT-RDPO/results_2024-10-09T01-39-59.424102.json,princeton-nlp/Mistral-7B-Base-SFT-RDPO,False,7241732096,0.8908058468228637,0.716600017001933,0.6615451571036659,0.7516824966868938 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Base-SFT-RRHF/results_2024-10-09T02-56-01.512965.json,princeton-nlp/Mistral-7B-Base-SFT-RRHF,False,7241732096,0.8417188384976202,0.688357345327554,0.446593203434771,0.6765878805563661 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Base-SFT-SimPO/results_2024-05-30T05-20-57.155719.json,princeton-nlp/Mistral-7B-Base-SFT-SimPO,False,7241732096,0.8874793724718586,0.7203452072610523,0.6109324432630949,0.8134649881602225 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Base-SFT-SimPO/results_2024-09-24T06-40-34.860235.json,princeton-nlp/Mistral-7B-Base-SFT-SimPO,False,7241732096,0.8925601781585555,0.6981678999006435,0.6642180413063468,0.7391020879888681 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Base-SFT-SimPO/results_2024-10-04T03-02-16.828129.json,princeton-nlp/Mistral-7B-Base-SFT-SimPO,False,7241732096,0.8925601781585555,0.6979065334676195,0.6642180413063468,0.7375983536256915 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Base-SFT-SLiC-HF/results_2024-10-09T04-37-54.415175.json,princeton-nlp/Mistral-7B-Base-SFT-SLiC-HF,False,7241732096,0.8925613256216668,0.7201250909754648,0.7181621092384083,0.7067782106289907 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Instruct-CPO/results_2024-10-09T06-01-15.185745.json,princeton-nlp/Mistral-7B-Instruct-CPO,False,7241732096,0.9056372391556865,0.7572882369786968,0.7060474998103043,0.8060416666666667 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Instruct-DPO/results_2024-10-09T07-29-34.620210.json,princeton-nlp/Mistral-7B-Instruct-DPO,False,7241732096,0.89733161586644,0.7643656679970449,0.7007313681868743,0.8200778971500587 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Instruct-IPO/results_2024-10-18T04-53-46.545059.json,princeton-nlp/Mistral-7B-Instruct-IPO,False,7241732096,0.8990256695860119,0.7597181962843274,0.7145366710584102,0.795430441898527 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Instruct-KTO/results_2024-10-10T01-39-05.388957.json,princeton-nlp/Mistral-7B-Instruct-KTO,False,7241732096,0.9031616961124372,0.7682182834894572,0.6762975530289601,0.796400920694706 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Instruct-ORPO/results_2024-10-10T03-21-24.342706.json,princeton-nlp/Mistral-7B-Instruct-ORPO,False,7241732096,0.9027775182021824,0.7380242782727071,0.6626835097302833,0.7670499708547507 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Instruct-RDPO/results_2024-10-10T04-57-07.262898.json,princeton-nlp/Mistral-7B-Instruct-RDPO,False,7241732096,0.9035715358705848,0.7598341554806352,0.6763551498057019,0.8191248121481978 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Instruct-RRHF/results_2024-10-10T06-56-37.421258.json,princeton-nlp/Mistral-7B-Instruct-RRHF,False,7241732096,0.9027712879720389,0.7395853027740446,0.6940669411726398,0.7663748206092372 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Instruct-SimPO/results_2024-09-24T07-56-57.653424.json,princeton-nlp/Mistral-7B-Instruct-SimPO,False,7241732096,0.8776596674356947,0.727318229917645,0.6655094956095569,0.773431614882243 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/princeton-nlp/Mistral-7B-Instruct-SLiC-HF/results_2024-10-10T08-27-01.104607.json,princeton-nlp/Mistral-7B-Instruct-SLiC-HF,False,7241732096,0.904411700902765,0.7498267129210576,0.6639545926205676,0.7830689803874399 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/prithivMLmods/Megatron-Opus-14B-Exp/results_2025-02-05T11-03-56.616036.json,prithivMLmods/Megatron-Opus-14B-Exp,False,14765947904,0.9456339276992554,0.8496917462691944,0.7867159436308373,0.8857612656789584 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/prithivMLmods/Qwen2.5-14B-DeepSeek-R1-1M/results_2025-01-31T23-33-58.338546.json,prithivMLmods/Qwen2.5-14B-DeepSeek-R1-1M,False,14770033664,0.8185192925422811,0.7986042203854202,0.516147503435639,0.8981315230669585 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/projecte-aina/aguila-7b/results_2024-02-08T03-52-57.903345.json,projecte-aina/aguila-7b,False,6854619456,0.3333333333333333,0.020311402918737537,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/projecte-aina/aguila-7b/results_2024-04-04T10-03-27.948314.json,projecte-aina/aguila-7b,False,6854619456,0.49279497098646036,0.07440697462061993,0.4396551724137931,0.3389990557129367 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/projecte-aina/FLOR-1.3B/results_2024-02-08T03-13-08.820840.json,projecte-aina/FLOR-1.3B,False,1311533056,0.3333333333333333,0.09691690000196435,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/projecte-aina/FLOR-1.3B/results_2024-04-04T08-56-25.624874.json,projecte-aina/FLOR-1.3B,False,1311533056,0.345026924452679,0.015562148096236503,0.4396551724137931,0.395407721696143 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/projecte-aina/FLOR-6.3B/results_2024-02-08T03-29-01.274461.json,projecte-aina/FLOR-6.3B,False,6247264256,0.4438707294213775,0.0070054014064472835,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/projecte-aina/FLOR-6.3B/results_2024-04-04T09-51-36.124295.json,projecte-aina/FLOR-6.3B,False,6247264256,0.44539372798813676,0.012774037998423506,0.4396551724137931,0.35671043584712653 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/projecte-aina/FLOR-760M/results_2024-02-08T02-51-18.675380.json,projecte-aina/FLOR-760M,False,757157376,0.3333333333333333,0.007895261532889483,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/projecte-aina/FLOR-760M/results_2024-04-04T08-06-22.356408.json,projecte-aina/FLOR-760M,False,757157376,0.41180825593044773,0.07618967054960082,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/proxectonos/Cerebras-1.3B-GL/results_2024-04-24T12-55-56.258777.json,proxectonos/Cerebras-1.3B-GL,False,1315723264,0.3333333333333333,0.1998219477433199,0.4396551724137931,0.36520703713341696 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/proxectonos/Cerebras-1.3B-GL/results_2024-04-28T12-04-16.087879.json,proxectonos/Cerebras-1.3B-GL,False,1315723264,0.3333333333333333,0.14846679911482175,0.4396551724137931,0.3847087284239676 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/proxectonos/FLOR-1.3B-GL/results_2024-04-24T12-41-22.775798.json,proxectonos/FLOR-1.3B-GL,False,1311533056,0.3582588385476761,0.034121946121581384,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/proxectonos/FLOR-1.3B-GL/results_2024-04-28T08-11-09.050685.json,proxectonos/FLOR-1.3B-GL,False,1311533056,0.34680711177144763,0.0018817748209403618,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/proxectonos/Llama-3.1-Carballo/results_2024-10-26T03-00-46.269684.json,proxectonos/Llama-3.1-Carballo,False,8030261248,0.5312299465240642,0.20132833844993506,0.5240274599542334,0.6307471007581653 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/proxectonos/Llama-3.1-Carballo/results_2025-04-05T22-30-40.653727.json,proxectonos/Llama-3.1-Carballo,False,8030261248,0.5477679046789041,0.16028198081637565,0.5308270676691729,0.6576982621318492 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/PrunaAI/dbrx-base-bnb-4bit/results_2024-05-02T01-12-55.471013.json,PrunaAI/dbrx-base-bnb-4bit,False,131596523520,0.9280791802073519,0.7917361086673437,0.45449901481427424,0.6659313413109303 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/PrunaAI/dbrx-instruct-bnb-4bit/results_2024-05-18T10-08-36.556587.json,PrunaAI/dbrx-instruct-bnb-4bit,False,131596523520,0.9297385620915033,0.8334641713026693,0.6141119512543087,0.7259225100277558 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/PrunaAI/Jamba-v0.1-bnb-4bit/results_2024-05-01T21-10-51.867015.json,PrunaAI/Jamba-v0.1-bnb-4bit,False,51570323328,0.8443951572017836,0.7346843581320739,0.6238473593282052,0.7456172415930774 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/pucpr/gpt2-bio-pt/results_2024-02-07T21-28-38.760468.json,pucpr/gpt2-bio-pt,False,124439808,0.22210105416212286,0.012461962425822393,0.08917890157694398,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/pucpr/gpt2-bio-pt/results_2024-03-08T03-04-30.604202.json,pucpr/gpt2-bio-pt,False,124439808,0.11320420958761174,0.03476467293472095,0.19574809466506216,0.24738051988698073 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/qihoo360/TinyR1-32B-Preview/results_2025-07-10T05-50-00.414573.json,qihoo360/TinyR1-32B-Preview,False,32763876352,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/qnguyen3/Master-Yi-9B/results_2024-06-12T10-51-12.606235.json,qnguyen3/Master-Yi-9B,False,8829407232,0.8868232701808685,0.7470955119173577,0.649412506638343,0.8674968831071828 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-14B/results_2024-03-09T09-10-27.408577.json,Qwen/Qwen-14B,False,14167290880,0.9060777099924455,0.7697342708611143,0.5204929592744473,0.807940417546329 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-14B/results_2024-04-20T16-52-41.258343.json,Qwen/Qwen-14B,False,14167290880,0.905257427983646,0.7702794193742895,0.5163130653349429,0.8051893701611216 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-1_8B/results_2024-02-07T11-00-37.396760.json,Qwen/Qwen-1_8B,False,1836828672,0.3547570850202429,0.02235404928442581,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-1_8B/results_2024-02-17T08-14-45.289722.json,Qwen/Qwen-1_8B,False,1836828672,0.6483114247757344,0.1952565561550447,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-1_8B/results_2024-04-20T20-55-10.036369.json,Qwen/Qwen-1_8B,False,1836828672,0.647723552936823,0.20825159514319308,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-1_8B-Chat/results_2024-02-23T15-51-23.035475.json,Qwen/Qwen-1_8B-Chat,False,1836828672,0.6805974700732741,0.2428472411987268,0.43614248479582973,0.6560055396510498 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-1_8B-Chat/results_2024-04-20T18-02-22.288187.json,Qwen/Qwen-1_8B-Chat,False,1836828672,0.6789403882059499,0.2482452669328401,0.43652173913043485,0.6567993003026897 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-72B/results_2024-02-11T13-30-16.052007.json,Qwen/Qwen-72B,False,72287920128,0.8843159546229062,0.6681993412313705,0.7699555733206758,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-72B/results_2024-02-19T02-18-45.116973.json,Qwen/Qwen-72B,False,72287920128,0.9287560338396796,0.8003393548769813,0.7839813891658358,0.8763515067862895 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-72B-Chat/results_2024-03-04T09-41-26.834248.json,Qwen/Qwen-72B-Chat,False,72287920128,0.6781400966183575,0.0,0.10463621984974299,0.8440869615900215 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-7B/results_2024-02-09T20-57-09.165430.json,Qwen/Qwen-7B,False,7721324544,0.8091756669435983,0.24412106004029266,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-7B/results_2024-02-17T08-43-57.342331.json,Qwen/Qwen-7B,False,7721324544,0.8472520703047123,0.6858024275313188,0.4471267110923455,0.5893817694135484 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-7B/results_2024-04-20T22-33-41.771237.json,Qwen/Qwen-7B,False,7721324544,0.8459955556908465,0.684084749467917,0.4471267110923455,0.5914851384809467 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-7B-Chat/results_2024-02-26T12-26-15.833520.json,Qwen/Qwen-7B-Chat,False,7721324544,0.8549690332636068,0.6908122993427773,0.4844989185291997,0.7134974533106961 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen-7B-Chat/results_2024-04-20T21-21-50.272837.json,Qwen/Qwen-7B-Chat,False,7721324544,0.8537165389952162,0.6903948401756923,0.48898756788856057,0.7171411593106802 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-0.5B/results_2024-02-24T20-35-38.476499.json,Qwen/Qwen1.5-0.5B,False,463987712,0.4004081218058632,0.15331435595983967,0.5253674727358938,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-0.5B-Chat/results_2024-02-25T10-26-08.663727.json,Qwen/Qwen1.5-0.5B-Chat,False,463987712,0.3333333333333333,0.06880227241486703,0.1826813595904991,0.33427893671851205 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-1.8B/results_2024-02-22T17-34-27.100441.json,Qwen/Qwen1.5-1.8B,False,1836828672,0.42057189695322883,0.32199828051873347,0.19973603600941603,0.4317428166587821 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-1.8B-Chat/results_2024-02-23T15-22-26.031189.json,Qwen/Qwen1.5-1.8B-Chat,False,1836828672,0.7389755107329503,0.2582632571238322,0.3269443631144088,0.6703427485364883 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-110B/results_2024-08-13T00-47-28.114784.json,Qwen/Qwen1.5-110B,False,111209914368,0.8952178411699083,0.7916348433066385,0.8185224212621474,0.8387375870039644 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-110B-Chat/results_2024-05-24T17-17-26.943633.json,Qwen/Qwen1.5-110B-Chat,False,111209914368,0.9238241568808003,0.7616643629663564,0.8072566392674566,0.8397521389256764 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-110B-Chat/results_2024-05-26T01-01-39.902132.json,Qwen/Qwen1.5-110B-Chat,False,111209914368,0.9200225041420388,0.7541446204629966,0.7928832962550982,0.8489571799283399 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-14B/results_2024-02-23T05-16-02.933985.json,Qwen/Qwen1.5-14B,False,14167290880,0.9075315471912553,0.751878207925627,0.7475883440795721,0.8747463613817982 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-14B-Chat/results_2024-02-23T13-22-14.833209.json,Qwen/Qwen1.5-14B-Chat,False,14167290880,0.9321888182837614,0.780879738931287,0.7885841836734694,0.8329156223893066 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-32B/results_2024-04-13T19-30-56.827110.json,Qwen/Qwen1.5-32B,False,32512218112,0.9280834059429739,0.5301047372239611,0.7256362507759155,0.7760546898020694 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-32B-Chat/results_2024-04-14T04-34-10.988929.json,Qwen/Qwen1.5-32B-Chat,False,32512218112,0.9390883593858939,0.7567199748875164,0.7667587392454499,0.8098224152553263 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-4B/results_2024-02-22T20-25-32.369215.json,Qwen/Qwen1.5-4B,False,3950369280,0.8533420030265173,0.556935351593212,0.5275390900126778,0.5673638590540795 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-4B-Chat/results_2024-02-23T14-38-32.754712.json,Qwen/Qwen1.5-4B-Chat,False,3950369280,0.8411111832284026,0.7232382659994563,0.5705852196753889,0.8192848844714082 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-72B/results_2024-02-19T17-53-35.586754.json,Qwen/Qwen1.5-72B,False,72287920128,0.9188128244863994,0.7167709775420129,0.7593729655207628,0.8761395878040701 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-72B-Chat/results_2024-02-20T12-32-34.348662.json,Qwen/Qwen1.5-72B-Chat,False,72287920128,0.927959115835123,0.7820328772395573,0.7948250481593255,0.8684077335762729 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-7B/results_2024-02-23T01-12-01.874268.json,Qwen/Qwen1.5-7B,False,7721324544,0.8944172809654745,0.687082724187013,0.48994341882459835,0.8150798965430385 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-7B-Chat/results_2024-02-26T11-34-46.128144.json,Qwen/Qwen1.5-7B-Chat,False,7721324544,0.8924941577812342,0.7011640993361781,0.7245623355596034,0.7810188120904671 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-MoE-A2.7B/results_2024-04-17T03-18-25.644273.json,Qwen/Qwen1.5-MoE-A2.7B,False,14315784192,0.8433996114156894,0.3631361487004874,0.4396551724137931,0.3380776303748972 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen1.5-MoE-A2.7B-Chat/results_2024-04-17T20-22-30.814298.json,Qwen/Qwen1.5-MoE-A2.7B-Chat,False,14315784192,0.8973604280190155,0.6905476485946694,0.6393020594965675,0.6472125041810177 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2-0.5B/results_2024-06-12T06-46-45.592981.json,Qwen/Qwen2-0.5B,False,494032768,0.5874337950921603,0.02509703564375903,0.43868739205526774,0.4634868211245766 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2-0.5B-Instruct/results_2024-06-12T08-50-35.262249.json,Qwen/Qwen2-0.5B-Instruct,False,494032768,0.4010616639567136,0.5203075242304281,0.5734565832149147,0.7233748271092669 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2-57B-A14B/results_2024-08-14T23-15-57.325481.json,Qwen/Qwen2-57B-A14B,False,57408658944,0.9443908461902331,0.7807805618732349,0.5384615384615385,0.7257525480675053 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2-72B/results_2024-07-06T22-46-30.899100.json,Qwen/Qwen2-72B,False,72706203648,0.9332353396385054,0.8335634101713818,0.7636129202896155,0.7736575121214886 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2-72B-Instruct/results_2024-06-14T13-11-55.047025.json,Qwen/Qwen2-72B-Instruct,False,72706203648,0.9493366322025627,0.7472406013787171,0.8439608636977058,0.8393795461324889 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2-72B-Instruct/results_2024-08-13T20-10-21.334329.json,Qwen/Qwen2-72B-Instruct,False,72706203648,0.9468894337609904,0.7511895812594225,0.8494905835134301,0.843196417585131 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2-7B/results_2024-06-12T05-41-22.096154.json,Qwen/Qwen2-7B,False,7615616512,0.9121875863252722,0.7143334641279073,0.6229260935143288,0.8231060017606546 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2-7B-Instruct/results_2024-06-12T05-41-21.924329.json,Qwen/Qwen2-7B-Instruct,False,7615616512,0.9391254439562453,0.7574397823735723,0.7590078425046203,0.7862560415950182 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-0.5B/results_2024-12-04T00-57-38.844522.json,Qwen/Qwen2.5-0.5B,False,494032768,0.5654469172848768,0.00446731205420841,0.4396551724137931,0.4378808032996366 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-0.5B-Instruct/results_2024-12-04T05-00-51.016352.json,Qwen/Qwen2.5-0.5B-Instruct,False,494032768,0.8361895029238644,0.547593093652944,0.5988445062182991,0.5089347433800673 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-1.5B/results_2024-12-03T21-41-04.719094.json,Qwen/Qwen2.5-1.5B,False,1543714304,0.8121131741821397,0.6040839339033337,0.4396551724137931,0.4711821853167551 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-1.5B-Instruct/results_2024-12-04T04-27-36.831054.json,Qwen/Qwen2.5-1.5B-Instruct,False,1543714304,0.8776976565681645,0.7563027241318341,0.7257942232409678,0.7150380139289418 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-14B/results_2024-10-14T01-39-05.879176.json,Qwen/Qwen2.5-14B,False,14770033664,0.9386663744745218,0.8519686617050359,0.834501651760535,0.7965063957756449 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-14B/results_2025-07-07T13-08-28.749937.json,Qwen/Qwen2.5-14B,False,14770033664,0.9390737082987084,0.8230338200699858,0.8386914833615341,0.793466419918363 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-14B-Instruct/results_2024-12-04T09-29-07.819263.json,Qwen/Qwen2.5-14B-Instruct,False,14770033664,0.9477121041913218,0.8456831074944621,0.8009931245225363,0.8902087417829547 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-14B-Instruct/results_2024-12-07T05-12-38.832139.json,Qwen/Qwen2.5-14B-Instruct,False,14770033664,0.9477121041913218,0.8456831074944621,0.8009931245225363,0.8902087417829547 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-14B-Instruct/results_2025-07-07T08-12-11.339710.json,Qwen/Qwen2.5-14B-Instruct,False,14770033664,0.9473037017333983,0.8049005208176889,0.7978493490063782,0.890936897154724 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-14B-Instruct-1M/results_2025-02-05T12-30-01.262130.json,Qwen/Qwen2.5-14B-Instruct-1M,False,14770033664,0.9472792094018343,0.8131303910610806,0.7782478165938864,0.9135392704735588 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-32B/results_2024-12-02T21-59-10.647096.json,Qwen/Qwen2.5-32B,False,32763876352,0.9443789544590661,0.8361049732899307,0.8685587583148558,0.885055347902568 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-32B-Instruct/results_2024-12-02T19-10-06.084404.json,Qwen/Qwen2.5-32B-Instruct,False,32763876352,0.9472943438455241,0.7888206059917943,0.8226761081260601,0.9191698715148393 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-3B/results_2024-12-03T19-39-25.688292.json,Qwen/Qwen2.5-3B,False,3085938688,0.9098765622165653,0.6945754782265189,0.5314377252703244,0.8059734527003406 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-3B-Instruct/results_2024-12-04T03-45-15.186841.json,Qwen/Qwen2.5-3B-Instruct,False,3085938688,0.9219713820972144,0.7915221440649431,0.7709654686398872,0.7667810479361502 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-72B-Instruct/results_2024-12-05T08-33-20.190072.json,Qwen/Qwen2.5-72B-Instruct,False,72706203648,0.9513853113589974,0.8024589729352112,0.8178353231321114,0.8802651889069097 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-7B/results_2024-10-16T02-59-27.647376.json,Qwen/Qwen2.5-7B,False,7615616512,0.9354523031530886,0.7706906472946917,0.5749166836699993,0.6851397629607747 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-7B-Instruct/results_2024-10-16T01-59-10.129793.json,Qwen/Qwen2.5-7B-Instruct,False,7615616512,0.938724467693844,0.7618358910352698,0.8304065025676917,0.7798498552716642 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-7B-Instruct/results_2025-05-05T08-05-41.029779.json,Qwen/Qwen2.5-7B-Instruct,False,7615616512,0.9383164891020759,0.7611102067661989,0.8280825203902127,0.7798498552716642 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-7B-Instruct-1M/results_2025-02-01T00-59-15.699787.json,Qwen/Qwen2.5-7B-Instruct-1M,False,7615616512,0.9317750232595009,0.8276681214676489,0.8159161710563579,0.7714640721552268 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-Coder-14B/results_2024-11-14T01-35-28.892415.json,Qwen/Qwen2.5-Coder-14B,False,14770033664,0.9337753918108967,0.8133069333318578,0.7983969057117837,0.7512986514633611 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-Coder-14B-Instruct/results_2024-11-14T06-51-08.271124.json,Qwen/Qwen2.5-Coder-14B-Instruct,False,14770033664,0.9407479519433439,0.8383279450252241,0.7670573030568968,0.7300343567946332 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-Coder-32B/results_2024-12-09T00-13-52.812489.json,Qwen/Qwen2.5-Coder-32B,False,32763876352,0.9357811302431517,0.7904533269627655,0.8329256012410378,0.6618403077365544 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-Coder-32B-Instruct/results_2024-12-05T04-17-37.012266.json,Qwen/Qwen2.5-Coder-32B-Instruct,False,32763876352,0.9321676285911923,0.7832873556151732,0.8086349495409364,0.7848550686238143 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-Coder-7B/results_2024-10-16T01-45-35.230716.json,Qwen/Qwen2.5-Coder-7B,False,7615616512,0.9222805719476176,0.7805771909763752,0.4605164949992536,0.7387485204443534 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-Coder-7B/results_2024-10-16T01-45-35.238341.json,Qwen/Qwen2.5-Coder-7B,False,7615616512,0.9222805719476176,0.7805771909763752,0.4605164949992536,0.7387485204443534 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-Coder-7B/results_2025-05-07T13-26-27.518962.json,Qwen/Qwen2.5-Coder-7B,False,0,0.8804120065796365,0.6313051067048794,0.5573664990421456,0.691765127858891 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen2.5-Coder-7B-Instruct/results_2024-10-13T09-00-17.303333.json,Qwen/Qwen2.5-Coder-7B-Instruct,False,7615616512,0.8980917721943948,0.816849270112908,0.8048406862745099,0.8751022869201948 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-0.6B/results_2025-05-04T13-49-51.763673.json,Qwen/Qwen3-0.6B,False,596049920,0.4566260035449901,0.012645465849267477,0.071172976454591,0.2443576541602649 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-0.6B/results_2025-05-05T18-09-57.373002.json,Qwen/Qwen3-0.6B,False,0,0.7900011415593701,0.1525464594019102,0.5535013345894175,0.6809091421225949 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-0.6B-Base/results_2025-05-04T11-43-57.611981.json,Qwen/Qwen3-0.6B-Base,False,596049920,0.3333333333333333,0.07261417775627484,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-1.7B/results_2025-05-04T18-04-13.251730.json,Qwen/Qwen3-1.7B,False,1720574976,0.20222274104456625,0.028349855580939942,0.03505864354920959,0.3333925742857688 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-1.7B/results_2025-05-05T19-16-46.115589.json,Qwen/Qwen3-1.7B,False,0,0.9104459511400973,0.579185401179295,0.7776791739464026,0.829328412695491 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-1.7B-Base/results_2025-05-04T15-53-05.065564.json,Qwen/Qwen3-1.7B-Base,False,1720574976,0.8662637545694478,0.5553899414554937,0.4396551724137931,0.5852449303159961 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-14B/results_2025-05-06T11-36-43.577193.json,Qwen/Qwen3-14B,False,0,0.9333269059212339,0.7603114374326059,0.858169598947266,0.8688339606790159 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-14B-Base/results_2025-05-06T12-52-44.685402.json,Qwen/Qwen3-14B-Base,False,0,0.9057011909541014,0.7604960340494649,0.7277995944140367,0.7698717416405159 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-30B-A3B/results_2025-05-06T15-35-28.886729.json,Qwen/Qwen3-30B-A3B,False,0,0.9291844686201536,0.7821710164219243,0.8484897385417003,0.8662006553284444 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-30B-A3B-Base/results_2025-05-06T18-05-17.916872.json,Qwen/Qwen3-30B-A3B-Base,False,0,0.9125385316646653,0.7211436038560135,0.676682152990149,0.7365037939200618 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-32B/results_2025-05-06T20-36-38.421969.json,Qwen/Qwen3-32B,False,0,0.9366616038015056,0.7624424898631255,0.8360960331258545,0.8818723085679395 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-4B/results_2025-05-04T22-34-40.122673.json,Qwen/Qwen3-4B,False,4022468096,0.1882430647291942,0.2271215827197601,0.085883698381523,0.08510601800788037 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-4B/results_2025-05-06T01-33-09.830634.json,Qwen/Qwen3-4B,False,0,0.9293226941767623,0.7022251363729565,0.8252071844911102,0.8756571895258026 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-4B-Base/results_2025-05-05T00-28-32.832902.json,Qwen/Qwen3-4B-Base,False,4022468096,0.9387251220992052,0.817412088871353,0.46895424836601307,0.7720463481008193 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-8B/results_2025-05-06T04-33-42.673475.json,Qwen/Qwen3-8B,False,0,0.9258625550587084,0.7553404511190756,0.8326366169877619,0.8652847981253078 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/Qwen3-8B-Base/results_2025-05-06T05-40-58.545228.json,Qwen/Qwen3-8B-Base,False,0,0.89764867679431,0.6969834071065018,0.7378419324116847,0.7802746645222898 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Qwen/QwQ-32B-Preview/results_2024-12-05T01-31-54.014053.json,Qwen/QwQ-32B-Preview,False,32763876352,0.9208176332204472,0.8308277837714224,0.8060518833151964,0.9206928808846877 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Ramikan-BR/Qwen2-0.5B-v5/results_2024-07-22T01-09-51.591921.json,Ramikan-BR/Qwen2-0.5B-v5,False,494032768,0.3333333333333333,0.024982214362657244,0.4396551724137931,0.5541090783965131 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Ramikan-BR/Qwen2-0.5B-v5/results_2024-07-22T01-09-51.592689.json,Ramikan-BR/Qwen2-0.5B-v5,False,494032768,0.3333333333333333,0.024982214362657244,0.4396551724137931,0.5541090783965131 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Ramikan-BR/Qwen2-0.5B-v5/results_2024-07-22T01-09-51.694824.json,Ramikan-BR/Qwen2-0.5B-v5,False,494032768,0.3333333333333333,0.024982214362657244,0.4396551724137931,0.5541090783965131 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Ramikan-BR/Qwen2-0.5B-v9/results_2025-02-04T08-06-56.544914.json,Ramikan-BR/Qwen2-0.5B-v9,False,494032768,0.3333333333333333,0.04347621081782021,0.4396551724137931,0.5229870877539271 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Ramikan-BR/TiamaPY-v30/results_2024-06-19T23-42-17.561440.json,Ramikan-BR/TiamaPY-v30,False,1100048384,0.3367698135882166,0.019616024795511965,0.4396551724137931,0.3389990557129367 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Ramikan-BR/TiamaPY-v31/results_2024-06-21T07-48-28.504692.json,Ramikan-BR/TiamaPY-v31,False,1100048384,0.33605087493777575,0.01935689428594359,0.4396551724137931,0.3389990557129367 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Ramikan-BR/tinyllama-coder-py-4bit-v10/results_2024-05-29T13-03-10.327337.json,Ramikan-BR/tinyllama-coder-py-4bit-v10,False,1100048384,0.4105763517528223,0.09183354777296178,0.45367132867132864,0.3558806058806059 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Ramikan-BR/tinyllama-coder-py-v11/results_2024-05-29T15-01-26.403443.json,Ramikan-BR/tinyllama-coder-py-v11,False,1100048384,0.4777644492911668,0.0465549816855676,0.4396551724137931,0.3622222119371426 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Ramikan-BR/tinyllama-coder-py-v12/results_2024-05-29T23-32-07.010938.json,Ramikan-BR/tinyllama-coder-py-v12,False,1100048384,0.3653844853319245,0.009190579858306076,0.5002423573852145,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/RDson/WomboCombo-R1-Coder-14B-Preview/results_2025-02-06T01-54-47.257414.json,RDson/WomboCombo-R1-Coder-14B-Preview,False,14770033664,0.9183734064113015,0.8445046642401395,0.6810071392410295,0.9228344656388816 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/bode-13b-alpaca-pt-br/results_2024-02-09T18-04-49.511422.json,recogna-nlp/bode-13b-alpaca-pt-br,False,13022417920,0.4260859463586567,0.37068485002111934,0.5754761980475085,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/bode-13b-alpaca-pt-br/results_2024-02-25T13-34-44.393221.json,recogna-nlp/bode-13b-alpaca-pt-br,False,13022417920,0.7122069215434327,0.4674730362338963,0.5168143799877225,0.8220549374629709 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/Bode-3.1-8B-Instruct-full/results_2024-09-01T01-33-24.575650.json,recogna-nlp/Bode-3.1-8B-Instruct-full,False,8030261248,0.9108047667597066,0.7622003162605613,0.7156998580615668,0.862024522051821 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/Bode-3.1-8B-Instruct-full/results_2024-09-01T02-41-26.513792.json,recogna-nlp/Bode-3.1-8B-Instruct-full,False,8030261248,0.9108047667597066,0.7622003162605613,0.7156998580615668,0.862024522051821 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/Bode-3.1-8B-Instruct-lora/results_2024-09-01T01-40-33.790084.json,recogna-nlp/Bode-3.1-8B-Instruct-lora,False,8084787200,0.9197286814442841,0.7429626239121664,0.6506364453818545,0.8840867693326709 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/bode-7b-alpaca-pt-br/results_2024-02-06T14-23-58.409626.json,recogna-nlp/bode-7b-alpaca-pt-br,False,6742609920,0.702446482303233,0.3367742885664205,0.6656319290465632,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/bode-7b-alpaca-pt-br/results_2024-02-17T00-17-03.515313.json,recogna-nlp/bode-7b-alpaca-pt-br,False,6742609920,0.7983437034736313,0.4346888805285222,0.6744821306624175,0.850631294632579 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/gembode-2b-base-ultraalpaca/results_2024-04-23T03-20-25.576942.json,recogna-nlp/gembode-2b-base-ultraalpaca,False,2506172416,0.6951065529414797,0.30312822596457845,0.555462171025527,0.5318344778243523 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/gembode-2b-base-ultraalpaca-qlora/results_2024-04-23T17-54-18.241577.json,recogna-nlp/gembode-2b-base-ultraalpaca-qlora,False,2584619008,0.6974955097348947,0.04162669962154587,0.5262883914627798,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/GemBode-2b-it/results_2024-04-01T21-06-43.218136.json,recogna-nlp/GemBode-2b-it,False,2506172416,0.5310352237070368,0.15565576448410878,0.5304956090884898,0.6689301037127124 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/GemBode-2b-it/results_2024-04-04T14-37-58.641441.json,recogna-nlp/GemBode-2b-it,False,2506172416,0.5270505348188489,0.1627527468902441,0.5294612794612795,0.675215194920858 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/gembode-2b-ultraalpaca/results_2024-04-11T23-40-55.082073.json,recogna-nlp/gembode-2b-ultraalpaca,False,2506172416,0.7131370059846825,0.34076850360449407,0.6008696229421137,0.47006437426999775 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/gembode-2b-ultraalpaca-qlora/results_2024-04-23T16-06-34.445180.json,recogna-nlp/gembode-2b-ultraalpaca-qlora,False,2584619008,0.3333333333333333,0.008744289928817567,0.4300399934413608,0.3640988372093023 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/gembode-7b/results_2024-04-24T01-26-52.427940.json,recogna-nlp/gembode-7b,False,8737696768,0.8660898922407284,0.7139025683213474,0.6739899260529418,0.798089617146851 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/gembode-7b-it/results_2024-04-24T03-48-11.198528.json,recogna-nlp/gembode-7b-it,False,8737696768,0.7909375144537008,0.6495212930759977,0.6467391304347826,0.8626528038653425 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/internlm-chatbode-20b/results_2024-05-15T13-57-12.679064.json,recogna-nlp/internlm-chatbode-20b,False,19861149696,0.9152793184707891,0.7894737787109948,0.8135882553772619,0.8171563741377617 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/internlm-chatbode-7b/results_2024-05-08T12-45-01.565674.json,recogna-nlp/internlm-chatbode-7b,False,7737708544,0.9132653061224489,0.8068834172948551,0.797966745710504,0.8798919027124412 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/internlm-chatbode2-7b-alpha/results_2024-06-12T02-37-01.198960.json,recogna-nlp/internlm-chatbode2-7b-alpha,False,8366428160,0.9154886524518426,0.8151071142980596,0.7908622908622909,0.858544742957652 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/internlm-chatbode2-7b-beta/results_2024-06-14T20-09-31.562507.json,recogna-nlp/internlm-chatbode2-7b-beta,False,8366428160,0.9162628416359759,0.8292073090106499,0.7798136131645113,0.8831607624963456 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/internlm2-chat-1_8b-ultracabrita/results_2024-05-06T02-44-19.717113.json,recogna-nlp/internlm2-chat-1_8b-ultracabrita,False,1889110016,0.8474467958203513,0.6031319002663305,0.4396551724137931,0.7230593961990576 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/internlmbode-7b/results_2024-04-05T12-56-22.889720.json,recogna-nlp/internlmbode-7b,False,7737708544,0.9074022870922758,0.8174268214865323,0.7538546753481643,0.8792807253769162 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/mistral-bode/results_2024-04-04T18-46-01.488696.json,recogna-nlp/mistral-bode,False,7262992384,0.856617431723123,0.6215348335578205,0.5644689097806366,0.732451433269857 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/mistralbode_7b_qlora_ultraalpaca/results_2024-04-23T20-19-47.179024.json,recogna-nlp/mistralbode_7b_qlora_ultraalpaca,False,7409504256,0.8891714080476152,0.7636816610855701,0.6716751326816636,0.8201601224933464 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/Phi-Bode/results_2024-04-02T01-30-46.274282.json,recogna-nlp/Phi-Bode,False,2779683840,0.6809918053051454,0.3056852557491311,0.4396551724137931,0.6050818260120585 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/phi-bode-2-ultraalpaca/results_2024-04-11T20-15-36.958650.json,recogna-nlp/phi-bode-2-ultraalpaca,False,2779683840,0.4538797946448635,0.24430661007443233,0.4396551724137931,0.541539447397188 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/phibode-3-mini-4k-ultraalpaca/results_2024-05-06T02-43-08.947553.json,recogna-nlp/phibode-3-mini-4k-ultraalpaca,False,3821079552,0.8855961737207682,0.6962827780132079,0.5064708415493756,0.8219118217312715 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/phibode_1_5_ultraalpaca/results_2024-05-16T03-57-14.492723.json,recogna-nlp/phibode_1_5_ultraalpaca,False,1418270720,0.6907406742960611,0.04935396761504174,0.4396551724137931,0.3493820791774778 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/phibode_1_5_ultraalpaca_qlora/results_2024-06-11T15-49-50.514379.json,recogna-nlp/phibode_1_5_ultraalpaca_qlora,False,1462310912,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/qwenbode_1_8b_chat_ultraalpaca/results_2024-04-23T04-39-57.242265.json,recogna-nlp/qwenbode_1_8b_chat_ultraalpaca,False,1836828672,0.6907365501362341,0.17891180047516525,0.4028734053736658,0.5287748382830149 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/qwenbode_1_8b_chat_ultraalpaca_qlora/results_2024-04-23T19-02-56.243195.json,recogna-nlp/qwenbode_1_8b_chat_ultraalpaca_qlora,False,1896794112,0.40523403335417163,0.04637904403463375,0.3214707901159244,0.6009601998164614 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recogna-nlp/zephyr_7b_beta_ultraalpaca/results_2024-04-23T20-20-06.853856.json,recogna-nlp/zephyr_7b_beta_ultraalpaca,False,7409504256,0.9068174032202809,0.6997490779884747,0.6514333951695879,0.8325186029114653 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recoilme/Gemma-2-Ataraxy-Gemmasutra-9B-slerp/results_2024-09-20T02-40-10.794476.json,recoilme/Gemma-2-Ataraxy-Gemmasutra-9B-slerp,False,9241705984,0.9320936326083178,0.8091252156880804,0.7738653343827351,0.8761395878040701 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recoilme/recoilme-gemma-2-9B-v0.2/results_2024-10-02T08-00-47.516085.json,recoilme/recoilme-gemma-2-9B-v0.2,False,9241705984,0.9366203005791933,0.7958273503268056,0.7868908109174575,0.8680218892032943 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recoilme/recoilme-gemma-2-9B-v0.3/results_2024-10-05T01-44-49.344787.json,recoilme/recoilme-gemma-2-9B-v0.3,False,9241705984,0.9386781739722916,0.797595648940614,0.7949795277090078,0.8455529664758781 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/recoilme/recoilme-gemma-2-9B-v0.4/results_2024-10-05T03-35-10.126788.json,recoilme/recoilme-gemma-2-9B-v0.4,False,9241705984,0.9345807442108435,0.7980167378081593,0.7150494520181769,0.8345817776577134 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Replete-AI/Replete-LLM-V2.5-Qwen-14b/results_2024-10-05T05-48-28.954342.json,Replete-AI/Replete-LLM-V2.5-Qwen-14b,False,14770033664,0.9468951058193111,0.8452503679397207,0.7994192015487959,0.88875166552086 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Replete-AI/Replete-LLM-V2.5-Qwen-7b/results_2024-10-05T06-21-40.443980.json,Replete-AI/Replete-LLM-V2.5-Qwen-7b,False,7615616512,0.9387253265980141,0.7582696305960869,0.8320806013113705,0.7830583484192929 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rfanucchi/Llama3-pt_br_cabrita/results_2024-05-21T18-03-08.788688.json,rfanucchi/Llama3-pt_br_cabrita,False,8084787200,0.8504302440518686,0.6905574565263299,0.4396551724137931,0.7725902207424897 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rfanucchi/Llama3-pt_br_cabrit_orca_qkvo/results_2024-06-12T09-20-06.634562.json,rfanucchi/Llama3-pt_br_cabrit_orca_qkvo,False,8084787200,0.8928472062085187,0.6390069397087196,0.4396551724137931,0.7914901797578582 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rfanucchi/Llama3-pt_br_orca/results_2024-05-18T04-17-48.547169.json,rfanucchi/Llama3-pt_br_orca,False,8084787200,0.8568464487644158,0.7296780837645185,0.4396551724137931,0.6562312949790503 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rfanucchi/Llama3-pt_bt_orca_cabrita_qkvLora/results_2024-12-07T07-43-11.181951.json,rfanucchi/Llama3-pt_bt_orca_cabrita_qkvLora,False,8068009984,0.877642825011246,0.6675379674224126,0.4396551724137931,0.7659466170873878 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/gemma-portuguese-2b-it/results_2024-04-11T20-12-36.339974.json,rhaymison/gemma-portuguese-2b-it,False,2506172416,0.2708413369189397,0.0,0.02659574468085106,0.2702883970489604 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/gemma-portuguese-luana-2b/results_2024-04-15T04-39-30.550543.json,rhaymison/gemma-portuguese-luana-2b,False,2506172416,0.7086250895822799,0.015059745163584971,0.4396551724137931,0.4004611738937028 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/gemma-portuguese-luana-2b/results_2024-04-19T03-27-12.107935.json,rhaymison/gemma-portuguese-luana-2b,False,2506172416,0.7088985088985089,0.013256112279022545,0.4396551724137931,0.39962372731297036 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/gemma-portuguese-tom-cat-2b-it/results_2024-04-30T13-01-42.167960.json,rhaymison/gemma-portuguese-tom-cat-2b-it,False,2506172416,0.7026723891391267,0.14057451236845117,0.4408902012248469,0.4659385571754847 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/Llama-3-portuguese-Tom-cat-8b-instruct/results_2024-05-01T02-33-12.225758.json,rhaymison/Llama-3-portuguese-Tom-cat-8b-instruct,False,8030261248,0.9090728393053975,0.7539877035655855,0.7605231647784839,0.8698937908496732 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/Llama-portuguese-13b-Luana-v0.2/results_2024-04-12T18-47-02.821845.json,rhaymison/Llama-portuguese-13b-Luana-v0.2,False,13015864320,0.658261903387606,0.4281293964356291,0.40437713546957244,0.8362279506681398 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/Llama3-portuguese-luana-8b-instruct/results_2024-04-25T13-49-49.822446.json,rhaymison/Llama3-portuguese-luana-8b-instruct,False,8030261248,0.8923524044881361,0.7286559455456432,0.6894295164709268,0.8592820512820513 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/Mistral-8x7b-portuguese-luana/results_2024-04-23T05-22-32.742075.json,rhaymison/Mistral-8x7b-portuguese-luana,False,46702792704,0.918207529097026,0.7994520046948086,0.7860385072937481,0.7241567974832157 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/Mistral-portuguese-luana-7b/results_2024-04-12T16-47-18.891306.json,rhaymison/Mistral-portuguese-luana-7b,False,7241732096,0.9062751880667184,0.7581000188868579,0.5778698664027709,0.7723894463024897 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/Mistral-portuguese-luana-7b-chat/results_2024-04-12T22-13-38.876769.json,rhaymison/Mistral-portuguese-luana-7b-chat,False,7241732096,0.9047401298302861,0.7654856046414533,0.6674701720310766,0.7745813653928725 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/Mistral-portuguese-luana-7b-Mathematics/results_2024-04-18T00-45-28.065413.json,rhaymison/Mistral-portuguese-luana-7b-Mathematics,False,7241732096,0.8935916363890852,0.7478455553944315,0.7486666666666666,0.763946859411585 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/Mistral-portuguese-luana-7b-Mathematics/results_2024-04-19T03-01-41.531579.json,rhaymison/Mistral-portuguese-luana-7b-Mathematics,False,7241732096,0.8906673114119923,0.7511980311776982,0.7346670193554287,0.7674448131094258 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/Mistral-portuguese-luana-7b-mental-health/results_2024-04-17T23-02-40.124756.json,rhaymison/Mistral-portuguese-luana-7b-mental-health,False,7241732096,0.9129607845788432,0.7497943690380084,0.6057101261727597,0.7686318178242204 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/Mistral-portuguese-luana-7b-mental-health/results_2024-04-19T01-54-55.276812.json,rhaymison/Mistral-portuguese-luana-7b-mental-health,False,7241732096,0.9117248747913189,0.751378985359473,0.6094454124857298,0.7743476941747574 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/phi-3-portuguese-tom-cat-4k-instruct/results_2024-04-30T10-46-11.430314.json,rhaymison/phi-3-portuguese-tom-cat-4k-instruct,False,3821079552,0.9154099953342754,0.7526854207444067,0.47460337127045094,0.8301138443791178 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/portuguese-phi1_5-tom-cat/results_2024-05-21T11-09-13.316485.json,rhaymison/portuguese-phi1_5-tom-cat,False,1418270720,0.3333333333333333,0.007954766700397426,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/portuguese-Phi3-Tom-Cat-128k-instruct/results_2024-05-23T22-41-55.737966.json,rhaymison/portuguese-Phi3-Tom-Cat-128k-instruct,False,3821079552,0.8885757978600564,0.6800200098531912,0.4515932999392298,0.8591649266539312 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/portuguese-tom-cat-13b/results_2024-05-20T17-41-34.035747.json,rhaymison/portuguese-tom-cat-13b,False,13015864320,0.7741212250576865,0.584353379958785,0.6814208185925483,0.841311660223144 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/Qwen-portuguese-luana-7b/results_2024-04-17T02-58-45.686566.json,rhaymison/Qwen-portuguese-luana-7b,False,7721324544,0.8105435578533237,0.7424537822607186,0.5796413196656436,0.7028522663288288 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhaymison/Qwen-portuguese-luana-7b/results_2024-04-19T11-14-40.377019.json,rhaymison/Qwen-portuguese-luana-7b,False,7721324544,0.8133352580295922,0.7376532197773,0.5833652227154447,0.7055255424464446 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhymes-ai/Aria/results_2024-12-06T02-16-03.328764.json,rhymes-ai/Aria,False,25307308912,0.900201705749847,0.7966754928142081,0.72497550029891,0.8488129543426489 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rhymes-ai/Aria-Base-64K/results_2024-12-06T17-14-07.499930.json,rhymes-ai/Aria-Base-64K,False,25307308912,0.8857939539659846,0.7865916195150257,0.6627665664806068,0.6424664840578924 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rishiraj/CatPPT/results_2024-02-26T19-43-35.008850.json,rishiraj/CatPPT,False,7241732096,0.9248100520349588,0.7800518971455536,0.7700800415702798,0.8553480360180306 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rishiraj/CatPPT-base/results_2024-02-26T18-41-56.705669.json,rishiraj/CatPPT-base,False,7241732096,0.9235804933406541,0.7784430066746557,0.769428955491387,0.8583629779928524 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/RLHFlow/ArmoRM-Llama3-8B-v0.1/results_2025-02-04T10-49-43.309390.json,RLHFlow/ArmoRM-Llama3-8B-v0.1,False,8030269440,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/RLHFlow/LLaMA3-iterative-DPO-final/results_2024-06-12T18-34-30.568502.json,RLHFlow/LLaMA3-iterative-DPO-final,False,8030261248,0.92278946649082,0.7087993617619036,0.6838748255929838,0.841362121955629 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/RogerioPiazzon/alpaca-lora-ptbr-7b/results_2024-05-27T15-51-04.203888.json,RogerioPiazzon/alpaca-lora-ptbr-7b,False,6742609920,0.3369541665038594,0.1817441507956813,0.5626911314984709,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rombodawg/Everyone-Coder-4x7b-Base/results_2024-07-28T02-42-40.107595.json,rombodawg/Everyone-Coder-4x7b-Base,False,24153690112,0.8949068924338127,0.6764417479682397,0.7038566806008666,0.8454215242053973 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rombodawg/Rombos-LLM-V2.5-Qwen-32b/results_2024-12-03T13-20-25.334575.json,rombodawg/Rombos-LLM-V2.5-Qwen-32b,False,32763876352,0.9468851789220416,0.7893155312064848,0.8226761081260601,0.9191609499354871 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rombodawg/Rombos-LLM-V2.6-Qwen-14b/results_2024-10-14T06-25-48.655425.json,rombodawg/Rombos-LLM-V2.6-Qwen-14b,False,14770033664,0.9468931561859193,0.8361919231611601,0.7937643178794838,0.9018302666807934 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rombodawg/rombos_Replete-Coder-Instruct-8b-Merged/results_2024-10-15T01-39-00.032530.json,rombodawg/rombos_Replete-Coder-Instruct-8b-Merged,False,8030261248,0.8613964104160183,0.7365266849498295,0.5863682644110275,0.8742649295087555 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/rombodawg/rombos_Replete-Coder-Llama3-8B/results_2024-10-18T06-12-41.069319.json,rombodawg/rombos_Replete-Coder-Llama3-8B,False,8030261248,0.8898165490147605,0.7403425083762252,0.5718323027559975,0.8380914618028019 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/royallab/ZephRP-m7b/results_2024-05-19T16-08-35.867510.json,royallab/ZephRP-m7b,False,7241732096,0.8915411564964209,0.72912193009527,0.6992338771035772,0.8014948945480438 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/RubielLabarta/LogoS-7Bx2-MoE-13B-v0.2/results_2024-07-04T01-29-22.119742.json,RubielLabarta/LogoS-7Bx2-MoE-13B-v0.2,False,12879138816,0.9264698029582955,0.7797640483645831,0.7802933673469388,0.7963636363636364 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Sakalti/Oxyge1-33B/results_2024-12-22T08-52-39.812103.json,Sakalti/Oxyge1-33B,False,32763876352,0.9468863490707113,0.7889847704512802,0.8249062367717624,0.9184409340659341 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/SakanaAI/DiscoPOP-zephyr-7b-gemma/results_2024-06-17T05-23-03.212449.json,SakanaAI/DiscoPOP-zephyr-7b-gemma,False,8537680896,0.848565239613503,0.7391207446164598,0.7263276079845404,0.8622880270505662 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Salesforce/SFR-Iterative-DPO-LLaMA-3-8B-R/results_2025-02-04T18-38-35.287392.json,Salesforce/SFR-Iterative-DPO-LLaMA-3-8B-R,False,8030261248,0.9191175930715705,0.6992380809613499,0.6987491957557714,0.8321754848998637 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/saltlux/luxia-21.4b-alignment-v1.0/results_2024-05-25T18-49-23.355684.json,saltlux/luxia-21.4b-alignment-v1.0,False,21421479936,0.9060276242466888,0.8119072027277104,0.6880394345731324,0.7294344250967385 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/saltlux/luxia-21.4b-alignment-v1.0/results_2024-05-25T20-54-11.729257.json,saltlux/luxia-21.4b-alignment-v1.0,False,21421479936,0.9043836192650903,0.8127132204493684,0.687367864693446,0.7311358109405715 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/saltlux/luxia-21.4b-alignment-v1.2/results_2024-06-12T23-35-46.058025.json,saltlux/luxia-21.4b-alignment-v1.2,False,21421479936,0.9015343943160568,0.7850553241810303,0.5814184936920059,0.7940479868190712 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Sao10K/L3-8B-Niitama-v1/results_2024-12-21T18-59-43.076987.json,Sao10K/L3-8B-Niitama-v1,False,8030261248,0.915212977332335,0.7283128221559086,0.7479154547217375,0.8565232719252938 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Sao10K/L3.1-8B-Niitama-v1.1/results_2025-02-04T22-02-55.741020.json,Sao10K/L3.1-8B-Niitama-v1.1,False,8030261248,0.9247750928387102,0.7521943306127783,0.7582702702702703,0.8606890863543011 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Saxo/Linkbricks-Horizon-AI-Avengers-V4-32B/results_2025-01-31T05-50-22.816002.json,Saxo/Linkbricks-Horizon-AI-Avengers-V4-32B,False,32763876352,0.9423637880119925,0.8321573127708237,0.8374505497154254,0.9256801592568016 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Saxo/Linkbricks-Horizon-AI-Avengers-V5-32B/results_2025-01-31T08-25-48.911527.json,Saxo/Linkbricks-Horizon-AI-Avengers-V5-32B,False,32763876352,0.9444110504676493,0.8324302282622531,0.8357894736842105,0.9242461367571629 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/SeaLLMs/SeaLLM-7B-v2/results_2024-08-08T05-46-17.568983.json,SeaLLMs/SeaLLM-7B-v2,False,7375949824,0.910123460086579,0.7586077181075638,0.7842076261469852,0.8023504608685286 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/SeaLLMs/SeaLLM-7B-v2.5/results_2024-05-18T14-04-41.229593.json,SeaLLMs/SeaLLM-7B-v2.5,False,8537680896,0.9151755082705118,0.7593949330856138,0.674173738017691,0.8903288138769379 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/SeaLLMs/SeaLLMs-v3-7B-Chat/results_2024-07-28T09-01-19.064982.json,SeaLLMs/SeaLLMs-v3-7B-Chat,False,7615616512,0.9284422748474763,0.7722497712304397,0.7992471385454668,0.73063484002582 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/semantixai/LloroV2/results_2024-04-02T11-32-45.449365.json,semantixai/LloroV2,False,6738546688,0.5719307475249892,0.2680507078742417,0.43771626297577854,0.6802387103939183 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/sethuiyer/Qwen2.5-7B-Anvita/results_2024-11-02T01-41-08.840143.json,sethuiyer/Qwen2.5-7B-Anvita,False,7615616512,0.9334056870125369,0.778663598103904,0.8097848807348216,0.7435781407906843 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/shadowml/BeagSake-7B/results_2024-06-15T19-11-48.353198.json,shadowml/BeagSake-7B,False,7241732096,0.6928997027815162,0.5649979800601793,0.5768283186744451,0.6577509670220215 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/shadowml/Mixolar-4x7b/results_2024-12-08T02-52-47.670165.json,shadowml/Mixolar-4x7b,False,36099461120,0.9293274119108301,0.8170748118090019,0.7581129587576001,0.7902382715971225 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/SinclairSchneider/zephyr-orpo-141b-A35b-v0.1-bnb-4bit/results_2024-05-26T17-04-50.909770.json,SinclairSchneider/zephyr-orpo-141b-A35b-v0.1-bnb-4bit,False,140620634112,0.9308991914086986,0.7976353266846197,0.8151483552943407,0.823619603839384 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Skywork/Skywork-13B-base/results_2024-02-13T13-13-14.633268.json,Skywork/Skywork-13B-base,False,13854113280,0.6640797445169774,0.109966284351211,0.5733609182848322,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Skywork/Skywork-13B-base/results_2024-04-02T19-24-01.591878.json,Skywork/Skywork-13B-base,False,13854113280,0.8716147105137804,0.39917156952926447,0.63055894421648,0.8085710379000774 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/sometimesanotion/Lamarck-14B-v0.6/results_2025-01-31T11-01-08.059808.json,sometimesanotion/Lamarck-14B-v0.6,False,14765947904,0.9427919759202827,0.8575907081913533,0.800014650941323,0.8923719379816355 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/sometimesanotion/Lamarck-14B-v0.7/results_2025-02-05T09-41-59.055000.json,sometimesanotion/Lamarck-14B-v0.7,False,14765947904,0.9419537895951781,0.8448669196781138,0.7988553117525761,0.9011372288949475 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/sometimesanotion/Qwen2.5-14B-Vimarckoso-v3/results_2025-01-31T12-31-27.244625.json,sometimesanotion/Qwen2.5-14B-Vimarckoso-v3,False,14765947904,0.944030110417436,0.8570969555169938,0.8059029395253282,0.8960083070408524 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/speakleash/Bielik-11B-v2/results_2024-10-17T01-39-05.112249.json,speakleash/Bielik-11B-v2,False,11168796672,0.9276352059225779,0.7799205784455716,0.6137142857142857,0.7404940125018982 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/speakleash/Bielik-11B-v2.0-Instruct/results_2024-10-18T02-03-08.462834.json,speakleash/Bielik-11B-v2.0-Instruct,False,11168796672,0.9259205634937818,0.801398553123075,0.8374505497154254,0.7788193164463637 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/speakleash/Bielik-11B-v2.0-Instruct/results_2024-10-18T02-07-51.708816.json,speakleash/Bielik-11B-v2.0-Instruct,False,11168796672,0.9259205634937818,0.801398553123075,0.8374505497154254,0.7788193164463637 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/speakleash/Bielik-11B-v2.1-Instruct/results_2024-10-18T01-47-30.900493.json,speakleash/Bielik-11B-v2.1-Instruct,False,11168796672,0.9350341785261549,0.7982096738474846,0.8273447309958466,0.7958148383005708 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/speakleash/Bielik-11B-v2.2-Instruct/results_2024-11-28T01-39-46.757551.json,speakleash/Bielik-11B-v2.2-Instruct,False,11168796672,0.936668538814805,0.8050395786153354,0.8031640750181548,0.7960372960372961 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/speakleash/Bielik-11B-v2.3-Instruct/results_2025-02-04T05-57-54.109014.json,speakleash/Bielik-11B-v2.3-Instruct,False,11168796672,0.9329452628161146,0.8048846756550186,0.8063301856457185,0.7958148383005708 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ssmits/Falcon2-5.5B-multilingual/results_2024-05-28T17-48-59.310480.json,ssmits/Falcon2-5.5B-multilingual,False,5465415680,0.0008156606851549756,0.0,0.04411764705882353,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ssmits/Falcon2-5.5B-Portuguese/results_2024-05-28T14-17-43.082234.json,ssmits/Falcon2-5.5B-Portuguese,False,5465415680,0.0008156606851549756,0.0,0.04411764705882353,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-2-12b/results_2024-04-13T11-23-48.720896.json,stabilityai/stablelm-2-12b,False,12143185920,0.8985556939105745,0.7289217128232647,0.7114058247557202,0.7623600870050067 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-2-12b-chat/results_2024-04-13T12-58-56.032815.json,stabilityai/stablelm-2-12b-chat,False,12143185920,0.9219432453832597,0.7405161976598901,0.6771141336487285,0.7989427397129121 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-2-1_6b/results_2024-02-07T02-19-52.381073.json,stabilityai/stablelm-2-1_6b,False,1644515328,0.42921528444709955,0.08322099881987416,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-2-1_6b/results_2024-03-08T07-07-06.260777.json,stabilityai/stablelm-2-1_6b,False,1644515328,0.33424046164366045,0.15788436345639384,0.4396551724137931,0.3396512838306731 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-2-1_6b/results_2024-04-20T05-21-43.409475.json,stabilityai/stablelm-2-1_6b,False,1644515328,0.33424046164366045,0.13007859321214196,0.4396551724137931,0.3349186726374015 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-2-1_6b-chat/results_2024-04-12T05-00-05.365629.json,stabilityai/stablelm-2-1_6b-chat,False,1644515328,0.6958297672377233,0.6285908805588103,0.446593203434771,0.5314433188988555 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-2-zephyr-1_6b/results_2024-03-08T08-53-16.434222.json,stabilityai/stablelm-2-zephyr-1_6b,False,1644515328,0.8651127286019933,0.6417654337468229,0.5099271173661724,0.7065528604613474 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-3b-4e1t/results_2024-02-07T02-35-22.857959.json,stabilityai/stablelm-3b-4e1t,False,2795443200,0.47137420034010846,0.006939634514205151,0.4471267110923455,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-3b-4e1t/results_2024-03-08T07-46-44.646469.json,stabilityai/stablelm-3b-4e1t,False,2795443200,0.48566367047977343,0.161620965224331,0.4396551724137931,0.6612083568605307 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-base-alpha-3b/results_2024-02-08T23-07-43.590555.json,stabilityai/stablelm-base-alpha-3b,False,3637321728,0.3333333333333333,0.027944186858664927,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-base-alpha-3b/results_2024-04-16T07-22-02.829511.json,stabilityai/stablelm-base-alpha-3b,False,3637321728,0.3333333333333333,0.05482636557770798,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-base-alpha-3b-v2/results_2024-02-08T21-44-32.013299.json,stabilityai/stablelm-base-alpha-3b-v2,False,2795934720,0.35299896655599283,0.021739027695983282,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-base-alpha-3b-v2/results_2024-04-16T03-05-11.469059.json,stabilityai/stablelm-base-alpha-3b-v2,False,2795934720,0.33875693585086425,0.15388088703910408,0.4396551724137931,0.43843615130743846 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-base-alpha-7b/results_2024-02-08T23-40-37.936580.json,stabilityai/stablelm-base-alpha-7b,False,7868755968,0.3333333333333333,0.008251040372066512,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-base-alpha-7b/results_2024-04-16T09-01-20.373870.json,stabilityai/stablelm-base-alpha-7b,False,7868755968,0.3442710085447083,0.07275995323484077,0.4396551724137931,0.3330157217722725 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-base-alpha-7b-v2/results_2024-02-08T22-34-53.379516.json,stabilityai/stablelm-base-alpha-7b-v2,False,6889414656,0.41477012011455616,0.2612742200838532,0.4678127512213221,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-base-alpha-7b-v2/results_2024-04-16T04-51-58.050879.json,stabilityai/stablelm-base-alpha-7b-v2,False,6889414656,0.5942379029702809,0.06250441072909244,0.5239722683331706,0.8012322375570973 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/stabilityai/stablelm-zephyr-3b/results_2024-03-08T08-57-43.114341.json,stabilityai/stablelm-zephyr-3b,False,2795443200,0.8756706040849712,0.6188996193211889,0.6709871216144451,0.6708045324270846 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/state-spaces/mamba-1.4b-hf/results_2024-04-18T23-42-19.165668.json,state-spaces/mamba-1.4b-hf,False,1372178432,0.3428635121469901,0.012831311142992859,0.4396551724137931,0.4675345929636195 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/state-spaces/mamba-130m-hf/results_2024-04-19T02-06-01.402679.json,state-spaces/mamba-130m-hf,False,129135360,0.543248630099413,0.03138351348073235,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/state-spaces/mamba-2.8b-hf/results_2024-04-11T20-03-36.310733.json,state-spaces/mamba-2.8b-hf,False,2768345600,0.5050802698742412,0.18075209102146808,0.4396551724137931,0.33710507019922026 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/state-spaces/mamba-370m-hf/results_2024-04-19T01-20-37.316572.json,state-spaces/mamba-370m-hf,False,371516416,0.5000531225430824,0.04039410635837621,0.4396551724137931,0.3322032156983613 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/state-spaces/mamba-790m-hf/results_2024-04-19T00-33-47.663320.json,state-spaces/mamba-790m-hf,False,793204224,0.5571495712779128,0.023278221074923514,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/t5-base/results_2024-02-14T03-48-55.037590.json,t5-base,False,222903552,0.21945826423438364,0.008119100920542854,0.11634819532908704,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/t5-base/results_2024-04-03T13-37-48.686073.json,t5-base,False,222903552,0.22120953194721352,0.03483992465952203,0.11785400175901495,0.22190293742017878 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/t5-large/results_2024-02-14T07-17-19.452890.json,t5-large,False,737668096,0.22304799920442644,0.028640141503653436,0.0948509485094851,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/t5-large/results_2024-04-05T01-05-32.137331.json,t5-large,False,737668096,0.1967674497012887,0.08766743905162708,0.10464426597217856,0.21837374237176313 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/t5-small/results_2024-02-14T01-51-06.060926.json,t5-small,False,60506624,0.2222222222222222,0.03024512521980755,0.09230769230769231,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/t5-small/results_2024-04-03T10-02-59.625167.json,t5-small,False,60506624,0.2206392694063927,0.008705938405819732,0.10127991096271564,0.22169059011164274 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tanliboy/lambda-qwen2.5-14b-dpo-test/results_2024-10-17T01-43-26.171758.json,tanliboy/lambda-qwen2.5-14b-dpo-test,False,14770033664,0.9448521957747049,0.8243398669298373,0.7882522522522523,0.9003808155770413 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tanliboy/lambda-qwen2.5-32b-dpo-test/results_2024-12-03T02-30-57.448037.json,tanliboy/lambda-qwen2.5-32b-dpo-test,False,32763876352,0.9464747005467062,0.786818651451759,0.8249062367717624,0.913360975307878 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TechxGenus/Jamba-v0.1-9B/results_2024-05-30T01-15-07.256137.json,TechxGenus/Jamba-v0.1-9B,False,9290690432,0.7516041341720514,0.12918337195290705,0.4396551724137931,0.452593537414966 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/teknium/OpenHermes-2-Mistral-7B/results_2024-02-27T00-53-05.335169.json,teknium/OpenHermes-2-Mistral-7B,False,7241748480,0.8986873329127992,0.6891008566875226,0.6618340732519423,0.7393515952707694 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/teknium/OpenHermes-2.5-Mistral-7B/results_2024-02-27T01-52-36.885316.json,teknium/OpenHermes-2.5-Mistral-7B,False,7241748480,0.9009851080543532,0.6887378174660355,0.7119751855544478,0.8017530796462489 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TencentARC/Mistral_Pro_8B_v0.1/results_2024-02-26T16-49-42.454942.json,TencentARC/Mistral_Pro_8B_v0.1,False,8986628096,0.8574374433475389,0.6850470140316613,0.4396551724137931,0.6885926784506906 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tensoropera/Fox-1-1.6B/results_2024-07-25T01-32-42.491874.json,tensoropera/Fox-1-1.6B,False,1665271808,0.7986102709656232,0.013767163804584731,0.4396551724137931,0.47240952642888057 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tensoropera/Fox-1-1.6B-Instruct-v0.1/results_2024-07-25T05-04-34.402031.json,tensoropera/Fox-1-1.6B-Instruct-v0.1,False,1665271808,0.5512347685522767,0.5372272064676139,0.6183024070348014,0.3349186726374015 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/terrycraddock/Reflection-Llama-3.1-8B/results_2024-09-19T06-17-56.815633.json,terrycraddock/Reflection-Llama-3.1-8B,False,8030261248,0.9030903950791537,0.7651789804795126,0.4396551724137931,0.7289174760362073 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TheBloke/wizardLM-7B-HF/results_2024-05-22T01-59-21.655252.json,TheBloke/wizardLM-7B-HF,False,6738423808,0.7985368487912935,0.3928416337999683,0.47714875915049165,0.6956506815169083 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TheBloke/zephyr-7B-beta-GPTQ/results_2024-05-29T07-56-34.401974.json,TheBloke/zephyr-7B-beta-GPTQ,False,1136201728,0.8747602497221476,0.6848780483036166,0.7380971253434792,0.8330124462550406 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TheDrummer/Big-Tiger-Gemma-27B-v1/results_2024-10-05T01-44-49.191653.json,TheDrummer/Big-Tiger-Gemma-27B-v1,False,27227128320,0.929232275191777,0.8127165758498651,0.7613044669772541,0.8849514433457764 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TheDrummer/Big-Tiger-Gemma-27B-v1/results_2025-07-07T14-08-09.344252.json,TheDrummer/Big-Tiger-Gemma-27B-v1,False,27227128320,0.586078506802503,0.31542295410004706,0.5457244258725787,0.6972731605905487 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TheDrummer/Gemmasutra-9B-v1/results_2025-02-04T23-05-39.687314.json,TheDrummer/Gemmasutra-9B-v1,False,9241705984,0.9300417017690956,0.8134611800233675,0.784170079361993,0.8620505357278279 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/agentlm-7b/results_2024-04-13T20-34-34.878222.json,THUDM/agentlm-7b,False,6740512768,0.7084033613445377,0.4390471903017725,0.5653909896663508,0.6302286113606869 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/chatglm2-6b/results_2024-04-13T23-58-57.021466.json,THUDM/chatglm2-6b,False,6243584000,0.727557280580944,0.5220053423222211,0.4808516661274668,0.5459280303030303 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/chatglm3-6b/results_2024-04-05T15-27-35.144602.json,THUDM/chatglm3-6b,False,6243584000,0.8311489805310066,0.578670198377361,0.6701812494834612,0.8026874828971815 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/chatglm3-6b-128k/results_2024-04-13T19-34-44.726389.json,THUDM/chatglm3-6b-128k,False,6243584000,0.8524446885244803,0.470502655043515,0.47388795045045046,0.7018017316691318 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/chatglm3-6b-32k/results_2024-04-05T17-22-01.113594.json,THUDM/chatglm3-6b-32k,False,6243584000,0.8859044964345155,0.7025258545086862,0.5985371800081267,0.7743565512987098 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/chatglm3-6b-base/results_2024-02-14T00-29-11.353268.json,THUDM/chatglm3-6b-base,False,6243584000,0.6004172619619615,0.7205871294107395,0.8115869651008071,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/chatglm3-6b-base/results_2024-04-03T08-21-29.467735.json,THUDM/chatglm3-6b-base,False,6243584000,0.9039224485412756,0.8004330870982324,0.8206498331000323,0.7754663188926945 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/GLM-4-32B-0414/results_2025-07-07T21-03-46.622283.json,THUDM/GLM-4-32B-0414,False,32566081536,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/glm-4-9b/results_2024-06-12T22-31-26.067481.json,THUDM/glm-4-9b,False,9399951360,0.9333736629407697,0.8128661201847864,0.8314315352697095,0.8110483364720653 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/GLM-4-9B-0414/results_2025-05-07T17-03-05.311978.json,THUDM/GLM-4-9B-0414,False,0,0.9379078336092272,0.7905105299669889,0.8685861538905879,0.8406112319446768 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/glm-4-9b-chat/results_2024-06-12T22-31-26.222149.json,THUDM/glm-4-9b-chat,False,9399951360,0.9313670058842967,0.8190999543668215,0.7782478165938864,0.8277717078541105 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/glm-4-9b-chat/results_2024-06-13T00-26-37.583558.json,THUDM/glm-4-9b-chat,False,9399951360,0.9325930787750805,0.8199633119800958,0.7775724600485918,0.8261628101105212 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/glm-4-9b-chat-1m/results_2024-06-12T23-29-01.907815.json,THUDM/glm-4-9b-chat-1m,False,9483857920,0.9379034815143636,0.8106987337312886,0.7626871120847024,0.822849314126908 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/glm-4-9b-chat-hf/results_2024-12-05T16-35-17.871895.json,THUDM/glm-4-9b-chat-hf,False,9399951360,0.9271802118629622,0.8107581674692308,0.757100149476831,0.8700378859564499 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/GLM-Z1-32B-0414/results_2025-07-07T23-17-30.039291.json,THUDM/GLM-Z1-32B-0414,False,32566081536,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/GLM-Z1-9B-0414/results_2025-05-07T17-21-54.047863.json,THUDM/GLM-Z1-9B-0414,False,0,0.9035790935916935,0.40324845882668453,0.7994664345294156,0.8796065470569719 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/GLM-Z1-Rumination-32B-0414/results_2025-07-15T01-29-13.399345.json,THUDM/GLM-Z1-Rumination-32B-0414,False,33141749760,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/LongAlign-7B-64k/results_2024-04-13T22-59-40.878063.json,THUDM/LongAlign-7B-64k,False,6740512768,0.7704193199567464,0.39121195804679265,0.48461310434746985,0.7779223622819595 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/THUDM/LongWriter-glm4-9b/results_2024-10-07T01-35-18.078840.json,THUDM/LongWriter-glm4-9b,False,9399951360,0.8402656499815521,0.7494676956507704,0.255058013895925,0.7495183221690935 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TIGER-Lab/AceCoder-Qwen2.5-7B-Ins-Rule/results_2025-04-07T09-08-13.246435.json,TIGER-Lab/AceCoder-Qwen2.5-7B-Ins-Rule,False,7615616512,0.9403591188432264,0.7598069041427369,0.8274180003951788,0.7790460211336232 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TIGER-Lab/AceCoder-Qwen2.5-Coder-7B-Base-Rule/results_2025-04-07T05-31-50.550452.json,TIGER-Lab/AceCoder-Qwen2.5-Coder-7B-Base-Rule,False,7615616512,0.9218916086702238,0.779649877895497,0.5378102624311676,0.7436259895137247 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TIGER-Lab/AceCoder-Qwen2.5-Coder-7B-Ins-Rule/results_2025-04-07T08-22-34.242574.json,TIGER-Lab/AceCoder-Qwen2.5-Coder-7B-Ins-Rule,False,7615616512,0.8964082358085583,0.8141587108786235,0.8071977469199038,0.8721786462426062 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TIGER-Lab/AceCodeRM-7B/results_2025-04-10T06-02-42.114156.json,TIGER-Lab/AceCodeRM-7B,False,7615616512,0.8963946654914912,0.8163579763711275,0.735485621269669,0.8655211824684415 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TIGER-Lab/MAmmoTH2-7B-Plus/results_2024-05-19T20-26-03.776683.json,TIGER-Lab/MAmmoTH2-7B-Plus,False,7241732096,0.8843035078991233,0.711905064393444,0.7423468521014304,0.6791294642857143 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TIGER-Lab/MAmmoTH2-8B-Plus/results_2024-05-19T20-28-01.440711.json,TIGER-Lab/MAmmoTH2-8B-Plus,False,8030261248,0.9104312286451928,0.673691215967068,0.6838748255929838,0.8799755052051439 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TIGER-Lab/Qwen2.5-32B-Instruct-CFT/results_2025-07-10T07-55-36.351327.json,TIGER-Lab/Qwen2.5-32B-Instruct-CFT,False,32763876352,0.9472931474849532,0.7802065781294906,0.8281603203317383,0.9198894801808208 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TIGER-Lab/Qwen2.5-Math-7B-CFT/results_2025-04-19T01-35-24.572463.json,TIGER-Lab/Qwen2.5-Math-7B-CFT,False,7615616512,0.8781683568800849,0.6870393006449449,0.45449901481427424,0.3818274700665496 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/falcon-11B/results_2024-05-18T15-03-05.414365.json,tiiuae/falcon-11B,False,11102822400,0.9064006761561062,0.7332618575347296,0.5182528935475422,0.7107889352593559 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/falcon-40b/results_2024-02-13T04-11-05.884844.json,tiiuae/falcon-40b,False,41303293952,0.4197802197802198,0.04513926929742694,0.4669333474933772,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/falcon-40b/results_2024-02-21T03-58-14.609880.json,tiiuae/falcon-40b,False,41303293952,0.8137537937671182,0.5632356011510816,0.4396551724137931,0.7601221082217126 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/falcon-7b/results_2024-02-09T19-33-31.081953.json,tiiuae/falcon-7b,False,6921720704,0.3333333333333333,0.06034092286086125,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/falcon-7b/results_2024-02-22T08-07-50.688709.json,tiiuae/falcon-7b,False,6921720704,0.4554241305414268,0.12397396118284368,0.4396551724137931,0.5431416664293376 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon-H1-0.5B-Base/results_2025-07-08T05-07-27.156126.json,tiiuae/Falcon-H1-0.5B-Base,False,521411104,0.35013262599469497,0.0,0.0,0.09789569990850869 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon-H1-0.5B-Instruct/results_2025-07-08T07-33-03.764027.json,tiiuae/Falcon-H1-0.5B-Instruct,False,521411104,0.37630258587417986,0.3394520329058093,0.18544935805991442,0.2641833810888252 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon-H1-1.5B-Deep-Base/results_2025-07-08T10-10-35.688671.json,tiiuae/Falcon-H1-1.5B-Deep-Base,False,1554872208,0.4517493271818531,0.0,0.26158940397350994,0.3839869281045752 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon-H1-1.5B-Deep-Instruct/results_2025-07-08T16-13-54.212155.json,tiiuae/Falcon-H1-1.5B-Deep-Instruct,False,1554874768,0.4394904458598726,0.7621449337692676,0.2840095465393795,0.43410852713178294 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon-H1-1.5B-Instruct/results_2025-07-08T22-50-47.544071.json,tiiuae/Falcon-H1-1.5B-Instruct,False,1554863488,0.42950391644908614,0.7260511850840119,0.2303370786516854,0.38747884940778343 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon-H1-34B-Base/results_2025-07-15T04-30-48.728728.json,tiiuae/Falcon-H1-34B-Base,False,33642516224,0.8793696984204424,0.0,0.0,0.8987797518590117 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon-H1-3B-Base/results_2025-07-09T04-02-33.419627.json,tiiuae/Falcon-H1-3B-Base,False,3149387264,0.454060236370568,0.0,0.0,0.3373052245646196 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon-H1-3B-Instruct/results_2025-07-09T09-36-26.386018.json,tiiuae/Falcon-H1-3B-Instruct,False,3149392384,0.4610517991300909,0.7846697811011769,0.34516129032258064,0.4284655300222387 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon-H1-7B-Base/results_2025-07-09T15-41-50.552612.json,tiiuae/Falcon-H1-7B-Base,False,7585648736,0.469908330011957,0.0,0.2878411910669975,0.36994727592267135 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon-H1-7B-Instruct/results_2025-07-09T21-17-40.440669.json,tiiuae/Falcon-H1-7B-Instruct,False,7585654880,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon3-10B-Base/results_2025-01-30T16-57-49.994219.json,tiiuae/Falcon3-10B-Base,False,10305653760,0.9205062830085595,0.7861299069660358,0.7543492775349265,0.6889672918269358 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon3-10B-Instruct/results_2025-01-30T20-31-36.005476.json,tiiuae/Falcon3-10B-Instruct,False,10305653760,0.9370673076923077,0.8009394707125536,0.8168992136555427,0.7936520263149136 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon3-1B-Base/results_2025-01-30T21-39-57.136674.json,tiiuae/Falcon3-1B-Base,False,1669408768,0.5430517362897619,0.1848122935552456,0.5005661598188289,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon3-1B-Instruct/results_2025-01-30T18-29-24.508725.json,tiiuae/Falcon3-1B-Instruct,False,1669408768,0.7607833655705996,0.21204486730376482,0.5195210827693909,0.6928264254996929 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon3-3B-Base/results_2025-01-30T22-12-44.476746.json,tiiuae/Falcon3-3B-Base,False,3227655168,0.7738274130153104,0.6179418737215836,0.5863636363636363,0.6402239278521884 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon3-3B-Instruct/results_2025-01-30T19-01-40.289408.json,tiiuae/Falcon3-3B-Instruct,False,3227655168,0.8740914856651618,0.6917542578575049,0.7001439309131616,0.7758265411925294 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon3-7B-Base/results_2025-01-29T04-05-34.417509.json,tiiuae/Falcon3-7B-Base,False,7455550464,0.9018326181978586,0.7850698923991428,0.71001221001221,0.7477831537431077 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tiiuae/Falcon3-7B-Instruct/results_2025-01-30T19-38-10.488929.json,tiiuae/Falcon3-7B-Instruct,False,7455550464,0.930888386729151,0.7719561347903747,0.7462677196306073,0.8572049130208099 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TinyLlama/TinyLlama-1.1B-Chat-v1.0/results_2024-02-24T15-52-37.358914.json,TinyLlama/TinyLlama-1.1B-Chat-v1.0,False,1100048384,0.4389940603879344,0.0052378590724758605,0.4396551724137931,0.3333333333333333 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T/results_2024-02-05T23-36-46.051635.json,TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T,False,1100048384,0.3333333333333333,0.051508356565052746,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T/results_2024-02-06T12-34-30.170523.json,TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T,False,1100048384,0.3333333333333333,0.051508356565052746,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T/results_2024-02-17T14-58-12.044782.json,TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T,False,1100048384,0.589309397637424,0.13574937116211067,0.4396551724137931,0.369192959530931 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TinyLlama/TinyLlama_v1.1/results_2024-07-26T05-59-47.399953.json,TinyLlama/TinyLlama_v1.1,False,1100048384,0.5412090194622737,0.1653828725329982,0.4396551724137931,0.34678939108409934 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TinyPixel/Llama-2-7B-bf16-sharded/results_2024-05-29T06-01-38.378610.json,TinyPixel/Llama-2-7B-bf16-sharded,False,6738415616,0.6353154692003844,0.3036032916828423,0.5303967191517392,0.7372127835097078 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tobiaspires/command-r-p11-fn/results_2024-07-05T17-47-15.000438.json,tobiaspires/command-r-p11-fn,False,34980831232,0.8803366916379576,0.6907322334378015,0.3618225084452432,0.9069232720458879 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/togethercomputer/RedPajama-INCITE-7B-Base/results_2024-02-07T06-08-50.798356.json,togethercomputer/RedPajama-INCITE-7B-Base,False,6857302016,0.3565103675260559,0.031663544686173564,0.4796980208120792,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/togethercomputer/RedPajama-INCITE-7B-Base/results_2024-04-02T05-35-09.270988.json,togethercomputer/RedPajama-INCITE-7B-Base,False,6857302016,0.46545879602571594,0.07300924060073302,0.47689087040026856,0.474037628651095 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/togethercomputer/RedPajama-INCITE-Base-3B-v1/results_2024-02-07T05-19-24.937552.json,togethercomputer/RedPajama-INCITE-Base-3B-v1,False,2775864320,0.3333333333333333,0.0025759159332480095,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/togethercomputer/RedPajama-INCITE-Base-3B-v1/results_2024-04-02T04-09-05.735716.json,togethercomputer/RedPajama-INCITE-Base-3B-v1,False,2775864320,0.4042526163839769,0.06694668265825117,0.4396551724137931,0.35714285714285715 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/tokyotech-llm/Llama-3-Swallow-8B-Instruct-v0.1/results_2024-09-13T01-40-33.534078.json,tokyotech-llm/Llama-3-Swallow-8B-Instruct-v0.1,False,8030261248,0.9037523178127644,0.7480754210337522,0.6487467314109804,0.8642812826541275 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TucanoBR/Tucano-160m/results_2024-12-04T14-57-09.634155.json,TucanoBR/Tucano-160m,False,162417408,0.33424046164366045,0.03604738980750616,0.4396551724137931,0.3557843229309216 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TucanoBR/Tucano-1b1/results_2024-12-04T14-08-15.254695.json,TucanoBR/Tucano-1b1,False,1100048384,0.5938355596818681,0.1456518796381127,0.4396551724137931,0.28616996597587385 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TucanoBR/Tucano-2b4/results_2024-12-04T13-38-16.578793.json,TucanoBR/Tucano-2b4,False,2444618240,0.5973672711217826,0.0,0.4396551724137931,0.3020560741553162 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TucanoBR/Tucano-2b4-Instruct/results_2024-12-04T12-06-51.875827.json,TucanoBR/Tucano-2b4-Instruct,False,2444628480,0.4331052025922113,0.012774037998423513,0.4396551724137931,0.4176001424952501 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/TucanoBR/Tucano-630m/results_2024-12-04T14-35-11.768529.json,TucanoBR/Tucano-630m,False,630253568,0.566662074665932,0.10904944101300668,0.4396551724137931,0.5278953302737802 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/UCLA-AGI/Gemma-2-9B-It-SPPO-Iter2/results_2024-08-09T02-40-40.201950.json,UCLA-AGI/Gemma-2-9B-It-SPPO-Iter2,False,9241705984,0.9407261843950693,0.782822567874957,0.7745519468704511,0.8765378928176637 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/UCLA-AGI/Gemma-2-9B-It-SPPO-Iter3/results_2024-08-09T01-40-37.088817.json,UCLA-AGI/Gemma-2-9B-It-SPPO-Iter3,False,9241705984,0.9436000072122752,0.7824894339212733,0.7768878718535469,0.8748846946385318 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/UCLA-AGI/Llama-3-Instruct-8B-SPPO-Iter2/results_2024-08-09T03-18-44.156930.json,UCLA-AGI/Llama-3-Instruct-8B-SPPO-Iter2,False,8030261248,0.9091166324509491,0.7141518081879549,0.6749819434413022,0.8639647822209131 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/UCLA-AGI/Llama-3-Instruct-8B-SPPO-Iter3/results_2024-08-09T04-19-33.155639.json,UCLA-AGI/Llama-3-Instruct-8B-SPPO-Iter3,False,8030261248,0.9112109957323393,0.7038791413087024,0.6519970018203234,0.8558459980359402 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/UCLA-AGI/Mistral7B-PairRM-SPPO/results_2024-10-02T07-04-29.768375.json,UCLA-AGI/Mistral7B-PairRM-SPPO,False,7241732096,0.8996550875340211,0.7177802175389065,0.6402927368130351,0.7805775801791581 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/UCLA-AGI/Mistral7B-PairRM-SPPO-Iter2/results_2024-08-09T04-26-37.476674.json,UCLA-AGI/Mistral7B-PairRM-SPPO-Iter2,False,7241732096,0.9056240104288098,0.7150956617943389,0.6549647590409369,0.7873748146886901 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/UCLA-AGI/Mistral7B-PairRM-SPPO-Iter3/results_2024-08-09T05-34-49.243967.json,UCLA-AGI/Mistral7B-PairRM-SPPO-Iter3,False,7241732096,0.9055449263660459,0.7078631967714201,0.6538474019827898,0.7827403231330002 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Unbabel/TowerBase-7B-v0.1/results_2024-02-06T23-18-53.380050.json,Unbabel/TowerBase-7B-v0.1,False,6738415616,0.424025213670537,0.09020044429155334,0.45217350015431734,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Unbabel/TowerBase-7B-v0.1/results_2024-02-17T05-36-36.290224.json,Unbabel/TowerBase-7B-v0.1,False,6738415616,0.3674233637116819,0.2067745699219024,0.446059558419109,0.6131741821396994 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unicamp-dl/ptt5-base-portuguese-vocab/results_2024-06-16T10-22-20.907058.json,unicamp-dl/ptt5-base-portuguese-vocab,False,222903552,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unicamp-dl/ptt5-small-portuguese-vocab/results_2024-06-13T03-00-33.208011.json,unicamp-dl/ptt5-small-portuguese-vocab,False,60506624,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unsloth/gemma-2b-bnb-4bit/results_2024-04-25T20-00-01.934859.json,unsloth/gemma-2b-bnb-4bit,False,2506172416,0.41771170006464126,0.3665030522390371,0.4439232761621561,0.6401737250423044 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unsloth/gemma-7b-bnb-4bit/results_2024-04-25T17-34-22.425788.json,unsloth/gemma-7b-bnb-4bit,False,8537680896,0.7755370609018697,0.5660376755750616,0.552788473326178,0.8316335486274552 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unsloth/llama-3-70b-bnb-4bit/results_2024-05-01T15-11-22.421210.json,unsloth/llama-3-70b-bnb-4bit,False,70553706496,0.8908379138287935,0.7477340698266566,0.4471267110923455,0.8438245629256864 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unsloth/llama-3-70b-Instruct-bnb-4bit/results_2024-05-01T08-03-17.411885.json,unsloth/llama-3-70b-Instruct-bnb-4bit,False,70553706496,0.9342311371831666,0.7852000713649874,0.8287377450980392,0.8840758598521377 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unsloth/llama-3-8b-bnb-4bit/results_2024-05-07T06-08-24.295588.json,unsloth/llama-3-8b-bnb-4bit,False,8030261248,0.8867343436157197,0.7688775235359817,0.4396551724137931,0.7353813925351123 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unsloth/llama-3-8b-Instruct-bnb-4bit/results_2024-04-25T00-37-14.806836.json,unsloth/llama-3-8b-Instruct-bnb-4bit,False,8030261248,0.8961900795070845,0.734368375585303,0.7618203309692672,0.8699320461302231 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unsloth/Llama-3.2-1B-Instruct/results_2024-11-25T01-39-13.774602.json,unsloth/Llama-3.2-1B-Instruct,False,1235814400,0.3852827239108828,0.19232706396526297,0.4959154280782978,0.34434915063412697 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unsloth/Llama-3.3-70B-Instruct-bnb-4bit/results_2024-12-20T23-50-37.361485.json,unsloth/Llama-3.3-70B-Instruct-bnb-4bit,False,70553706496,0.9248203427307905,0.7007451869593214,0.8285795431492677,0.8774263213918718 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unsloth/Meta-Llama-3.1-70B-bnb-4bit/results_2024-08-14T00-46-47.994707.json,unsloth/Meta-Llama-3.1-70B-bnb-4bit,False,70553706496,0.8960820177127455,0.7709568866116487,0.4396551724137931,0.8006416813281791 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unsloth/Meta-Llama-3.1-70B-Instruct-bnb-4bit/results_2024-08-14T02-21-18.748456.json,unsloth/Meta-Llama-3.1-70B-Instruct-bnb-4bit,False,70553706496,0.9248281216309403,0.6956846398113579,0.8310393407164702,0.8795664391810518 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unsloth/mistral-7b-bnb-4bit/results_2024-04-23T07-00-47.391024.json,unsloth/mistral-7b-bnb-4bit,False,7241732096,0.8867972721003294,0.6061626853956876,0.4471267110923455,0.7603043753963221 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unsloth/mistral-7b-v0.3-bnb-4bit/results_2024-05-23T00-49-42.217768.json,unsloth/mistral-7b-v0.3-bnb-4bit,False,7248023552,0.8704411371627762,0.5790158199742027,0.4396551724137931,0.7130292188431724 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/unsloth/Phi-3-mini-4k-instruct-bnb-4bit/results_2024-04-23T23-33-50.317982.json,unsloth/Phi-3-mini-4k-instruct-bnb-4bit,False,3821079552,0.8924533568486099,0.7079888468451605,0.6812356046329664,0.7960823376293866 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/upstage/SOLAR-10.7B-Instruct-v1.0/results_2024-02-24T23-56-01.386628.json,upstage/SOLAR-10.7B-Instruct-v1.0,False,10731524096,0.927692591384275,0.8166852379077367,0.7543037140136254,0.8039482041435217 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/upstage/SOLAR-10.7B-v1.0/results_2024-02-06T04-33-22.888391.json,upstage/SOLAR-10.7B-v1.0,False,10731524096,0.7143372127843074,0.7000542682426939,0.6469134202801768,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/upstage/SOLAR-10.7B-v1.0/results_2024-02-16T21-37-23.301583.json,upstage/SOLAR-10.7B-v1.0,False,10731524096,0.9185166017874254,0.7827021446757074,0.6517857142857143,0.8928551748909674 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/upstage/solar-pro-preview-instruct/results_2024-09-22T04-48-51.986065.json,upstage/solar-pro-preview-instruct,False,22140032000,0.9386663744745218,0.754051498045265,0.7752293057901838,0.8160470425400375 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/utter-project/EuroLLM-1.7B-Instruct/results_2025-04-07T04-54-54.471948.json,utter-project/EuroLLM-1.7B-Instruct,False,1656850432,0.6865462315167273,0.18935614285574542,0.18051871171646466,0.428351495982177 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/uukuguy/speechless-code-mistral-7b-v1.0/results_2024-06-16T05-21-00.580567.json,uukuguy/speechless-code-mistral-7b-v1.0,False,7241732096,0.912955321291761,0.7220708697514615,0.6011570827489481,0.7744650652492993 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/uukuguy/speechless-zephyr-code-functionary-7b/results_2024-07-28T07-44-33.861970.json,uukuguy/speechless-zephyr-code-functionary-7b,False,7241732096,0.8936596078263641,0.6876381793169524,0.5927601809954751,0.7757048748353097 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/uukuguy/zephyr-7b-alpha-dare-0.85/results_2024-05-19T02-44-11.587145.json,uukuguy/zephyr-7b-alpha-dare-0.85,False,7241732096,0.8787323773948668,0.6899792231838362,0.6731842114789772,0.8307008241707131 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/uygarkurt/llama-3-merged-linear/results_2024-05-18T12-22-49.038176.json,uygarkurt/llama-3-merged-linear,False,8030261248,0.9170332834196958,0.72913821581632,0.7494336310476937,0.8685110509928028 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/v000000/L3.1-Niitorm-8B-DPO-t0.0001/results_2024-10-05T05-22-41.006522.json,v000000/L3.1-Niitorm-8B-DPO-t0.0001,False,8030261248,0.921462472704961,0.7513514605933432,0.7567813392281477,0.8672714442455509 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/v000000/Qwen2.5-14B-Gutenberg-1e-Delta/results_2024-10-17T03-08-28.033489.json,v000000/Qwen2.5-14B-Gutenberg-1e-Delta,False,14770033664,0.946078431372549,0.8364823895770733,0.7968815254182839,0.8865641025641027 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/v000000/Qwen2.5-14B-Gutenberg-Instruct-Slerpeno/results_2024-10-03T01-41-21.029002.json,v000000/Qwen2.5-14B-Gutenberg-Instruct-Slerpeno,False,14770033664,0.9473037017333983,0.8430176000490963,0.7968815254182839,0.8880227399974159 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/v000000/Qwen2.5-Lumen-14B/results_2024-10-16T05-24-04.232475.json,v000000/Qwen2.5-Lumen-14B,False,14770033664,0.9464869191748698,0.8373451314325752,0.7953209554021272,0.8865641025641027 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/v000000/Qwen2.5-Lumen-14B/results_2024-12-07T06-43-02.205238.json,v000000/Qwen2.5-Lumen-14B,False,14770033664,0.9464869191748698,0.8373451314325752,0.7953209554021272,0.8865641025641027 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/v000000/Qwen2.5-Lumen-14B/results_2025-07-07T07-02-12.796608.json,v000000/Qwen2.5-Lumen-14B,False,14770033664,0.946078431372549,0.8028417333225462,0.7937643178794838,0.8880227399974159 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/v000000/SwallowMaid-8B-L3-SPPO-abliterated/results_2024-10-02T05-55-04.169030.json,v000000/SwallowMaid-8B-L3-SPPO-abliterated,False,8030261248,0.9127627391239099,0.7293756978101533,0.6843906848976986,0.8631774516440309 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/VAGOsolutions/Llama-3-SauerkrautLM-8b-Instruct/results_2024-05-19T22-55-30.334730.json,VAGOsolutions/Llama-3-SauerkrautLM-8b-Instruct,False,8030261248,0.9174464106844742,0.7270928056226034,0.7545670310701956,0.8663024455213106 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/VAGOsolutions/Llama-3.1-SauerkrautLM-8b-Instruct/results_2024-08-11T22-10-57.990812.json,VAGOsolutions/Llama-3.1-SauerkrautLM-8b-Instruct,False,8030261248,0.9239769323498082,0.7333969010282789,0.6928223763079412,0.871945221579223 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/VAGOsolutions/SauerkrautLM-Gemma-7b/results_2024-05-20T03-35-55.668254.json,VAGOsolutions/SauerkrautLM-Gemma-7b,False,8537680896,0.8423722782343472,0.6990587499891736,0.656952196937169,0.8607179487179487 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/VAGOsolutions/SauerkrautLM-Nemo-12b-Instruct/results_2024-08-07T05-12-30.136519.json,VAGOsolutions/SauerkrautLM-Nemo-12b-Instruct,False,12247782400,0.9264580217902157,0.759889315042445,0.8317563486813724,0.8197534327445837 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/VAGOsolutions/SauerkrautLM-Phi-3-medium/results_2025-02-05T07-27-43.561475.json,VAGOsolutions/SauerkrautLM-Phi-3-medium,False,13960238080,0.9267866138357781,0.7136951382155405,0.6947528802805143,0.8437238376960429 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/VAGOsolutions/SauerkrautLM-Qwen-32b/results_2024-06-14T02-39-11.342113.json,VAGOsolutions/SauerkrautLM-Qwen-32b,False,32512218112,0.4700047236655645,0.7499803526182416,0.39783300439038144,0.8189785637744593 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ValiantLabs/Llama3.1-8B-Cobalt/results_2024-10-05T08-11-08.878059.json,ValiantLabs/Llama3.1-8B-Cobalt,False,8030261248,0.9103201153952376,0.768279969478055,0.7932594329065015,0.8898480352486287 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ValiantLabs/Llama3.1-8B-Fireplace2/results_2024-08-11T22-23-28.466122.json,ValiantLabs/Llama3.1-8B-Fireplace2,False,8030261248,0.7845999749593089,0.6981551929380407,0.3623197356637742,0.8581081081081081 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/ValiantLabs/Llama3.1-8B-ShiningValiant2/results_2024-08-11T23-08-49.079492.json,ValiantLabs/Llama3.1-8B-ShiningValiant2,False,8030261248,0.9028821036720353,0.7432505806234806,0.7594940966732819,0.8664338662984536 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/CarbonBeagle-11B/results_2024-06-21T01-31-39.472689.json,vicgalle/CarbonBeagle-11B,False,10731524096,0.9210937556104944,0.8146595670108915,0.751597669118611,0.8482580645161291 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/CarbonBeagle-11B-truthy/results_2024-05-20T00-40-47.200785.json,vicgalle/CarbonBeagle-11B-truthy,False,10731524096,0.924836149775025,0.8120497093970859,0.7755116879339805,0.8180089679918182 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/Configurable-Hermes-2-Pro-Llama-3-8B/results_2024-06-15T11-57-14.697340.json,vicgalle/Configurable-Hermes-2-Pro-Llama-3-8B,False,8030523392,0.9223814244089994,0.7497178617324638,0.7416826178105915,0.8243026557887815 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/Configurable-Hermes-3-Llama-3.1-8B/results_2024-09-21T01-40-34.729149.json,vicgalle/Configurable-Hermes-3-Llama-3.1-8B,False,8030261248,0.8791258618533935,0.7144360325317087,0.727504719781441,0.8521907948281129 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/Configurable-Llama-3-8B-v0.3/results_2024-06-18T07-23-44.259829.json,vicgalle/Configurable-Llama-3-8B-v0.3,False,8030261248,0.919117161170951,0.758625910356018,0.7322061271238114,0.8603133994243684 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/Configurable-Llama-3.1-8B-Instruct/results_2024-08-11T15-33-58.941279.json,vicgalle/Configurable-Llama-3.1-8B-Instruct,False,8030261248,0.9190081749769909,0.7732640002400402,0.7002777777777778,0.8607142146501094 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/Configurable-Llama-3.1-8B-Instruct/results_2024-12-07T02-58-38.792305.json,vicgalle/Configurable-Llama-3.1-8B-Instruct,False,8030261248,0.9190081749769909,0.7732640002400402,0.7002777777777778,0.8607142146501094 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/Configurable-Mistral-7B/results_2024-06-15T11-08-28.661492.json,vicgalle/Configurable-Mistral-7B,False,7241732096,0.9088886719190068,0.7214777969410027,0.6821573456968995,0.783421311848054 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/Configurable-Yi-1.5-9B-Chat/results_2024-06-12T12-25-31.494281.json,vicgalle/Configurable-Yi-1.5-9B-Chat,False,8829407232,0.8749916559640499,0.7518470632490779,0.6776982003600871,0.8123256853529013 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/ConfigurableBeagle-11B/results_2024-06-15T01-31-31.325184.json,vicgalle/ConfigurableBeagle-11B,False,10731524096,0.9276957767969884,0.8158043682731361,0.7782188011607123,0.8290936613488302 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/ConfigurableBeagle-11B/results_2024-06-15T01-31-31.737501.json,vicgalle/ConfigurableBeagle-11B,False,10731524096,0.9276957767969884,0.8158043682731361,0.7782188011607123,0.8290936613488302 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/ConfigurableBeagle-11B/results_2024-06-16T01-31-29.633353.json,vicgalle/ConfigurableBeagle-11B,False,10731524096,0.9276957767969884,0.8158043682731361,0.7782188011607123,0.8290936613488302 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/ConfigurableBeagle-11B/results_2024-06-16T01-31-29.648962.json,vicgalle/ConfigurableBeagle-11B,False,10731524096,0.9276957767969884,0.8158043682731361,0.7782188011607123,0.8290936613488302 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/ConfigurableHermes-7B/results_2024-06-18T04-54-10.322985.json,vicgalle/ConfigurableHermes-7B,False,7241748480,0.8913326159687492,0.6521607883553826,0.7041461236714373,0.7655662127073287 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vicgalle/ConfigurableSOLAR-10.7B/results_2024-08-26T05-22-16.175643.json,vicgalle/ConfigurableSOLAR-10.7B,False,10731524096,0.9165101276141576,0.8271838931379497,0.7603967695452565,0.749459230696887 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/vilsonrodrigues/falcon-7b-instruct-sharded/results_2024-05-29T02-21-59.459463.json,vilsonrodrigues/falcon-7b-instruct-sharded,False,6921720704,0.5442507626719338,0.05084063725467101,0.4396551724137931,0.47451878248737356 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Walmart-the-bag/Misted-v2-7B/results_2024-04-18T23-28-15.566455.json,Walmart-the-bag/Misted-v2-7B,False,7241732096,0.9113400977866568,0.7517173052153636,0.6589071710738363,0.8156351971251167 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Walmart-the-bag/Quintellect-10.7B/results_2024-04-22T07-23-08.500560.json,Walmart-the-bag/Quintellect-10.7B,False,10731524096,0.9170701677123623,0.7627230529499934,0.7349043532671851,0.735005055611729 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Walmart-the-bag/WordWoven-2x7B/results_2024-04-18T23-29-01.257877.json,Walmart-the-bag/WordWoven-2x7B,False,12879138816,0.9092516992262643,0.7293099314404875,0.7500832099530581,0.8425736221065424 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/wandgibaut/periquito-3B/results_2024-02-16T13-25-42.275333.json,wandgibaut/periquito-3B,False,3554473600,0.38392470950636587,0.037061970342522815,0.4396551724137931,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/wandgibaut/periquito-3B/results_2024-03-08T02-05-32.490573.json,wandgibaut/periquito-3B,False,3554473600,0.4301327637723658,0.08919111846797594,0.4396551724137931,0.5046251022011318 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/wenbopan/Faro-Yi-34B/results_2024-06-15T17-52-54.255909.json,wenbopan/Faro-Yi-34B,False,34388917248,0.9157921173000471,0.7937406302545176,0.7184240249817289,0.8709784375416707 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/wenbopan/Faro-Yi-34B-DPO/results_2024-06-15T06-24-43.069723.json,wenbopan/Faro-Yi-34B-DPO,False,34388917248,0.9219611718575664,0.7977517918692778,0.7100198247711969,0.8512471655328798 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/wenbopan/Faro-Yi-9B/results_2024-06-14T13-02-23.332764.json,wenbopan/Faro-Yi-9B,False,8829407232,0.8838690279322889,0.7504692857083078,0.684183978132105,0.8192416060841443 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/wenbopan/Faro-Yi-9B-DPO/results_2024-06-15T01-29-02.867117.json,wenbopan/Faro-Yi-9B-DPO,False,8829407232,0.8852055030895261,0.720109675977982,0.7066189940969795,0.8102569662799524 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/wenbopan/Faro-Yi-9B-DPO/results_2024-06-18T13-15-41.173293.json,wenbopan/Faro-Yi-9B-DPO,False,8829407232,0.8852055030895261,0.720109675977982,0.7066189940969795,0.8102569662799524 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weni/WeniGPT-2.2.3-Zephyr-7B-LLM_Base_2.0.3_SFT/results_2024-04-13T02-20-29.545335.json,Weni/WeniGPT-2.2.3-Zephyr-7B-LLM_Base_2.0.3_SFT,False,7255363584,0.8700631503510818,0.28326240949151776,0.446593203434771,0.669133806441068 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weni/WeniGPT-2.2.3-Zephyr-7B-merged-LLM_Base_2.0.3_SFT/results_2024-04-13T19-09-39.898040.json,Weni/WeniGPT-2.2.3-Zephyr-7B-merged-LLM_Base_2.0.3_SFT,False,7241732096,0.869659474863291,0.2862752370355036,0.446593203434771,0.669133806441068 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weni/WeniGPT-2.4.1-Zephyr-7B-3-epochs-GPT-QA-1.0.1_DP_DPO/results_2024-04-14T03-35-32.685055.json,Weni/WeniGPT-2.4.1-Zephyr-7B-3-epochs-GPT-QA-1.0.1_DP_DPO,False,7248547840,0.8844680081937584,0.6873093329646505,0.6130518712248583,0.8071243382571215 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weni/WeniGPT-2.8.1-Zephyr-7B-zephyr-prompt-binarized/results_2024-04-14T07-02-23.873958.json,Weni/WeniGPT-2.8.1-Zephyr-7B-zephyr-prompt-binarized,False,7255363584,0.8848011534302536,0.33017790245015,0.45449901481427424,0.686159754526626 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weni/WeniGPT-2.8.1-Zephyr-7B-zephyr-prompt-binarized/results_2024-04-15T09-20-25.423963.json,Weni/WeniGPT-2.8.1-Zephyr-7B-zephyr-prompt-binarized,False,7255363584,0.8741641081717626,0.272697845605027,0.5674920634920635,0.504258658256007 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weni/WeniGPT-Agents-Zephyr-1.0.17-KTO/results_2024-04-15T05-29-30.182120.json,Weni/WeniGPT-Agents-Zephyr-1.0.17-KTO,False,7268995072,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weni/WeniGPT-Mistral-7B-instructBase/results_2024-04-13T05-37-06.823274.json,Weni/WeniGPT-Mistral-7B-instructBase,False,7268995072,0.6142066914966826,0.40022445504959214,0.344957407970794,0.6624304406677697 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weni/WeniGPT-Mistral-7B-instructBase-4bit/results_2024-04-14T01-06-41.169623.json,Weni/WeniGPT-Mistral-7B-instructBase-4bit,False,7268995072,0.7465864435011207,0.5521492658011754,0.3236178978633707,0.6902907646983476 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weni/ZeroShot-3.3.34-Mistral-7b-Multilanguage-3.3.0-merged/results_2024-04-14T08-35-42.065682.json,Weni/ZeroShot-3.3.34-Mistral-7b-Multilanguage-3.3.0-merged,False,7241732096,0.8937890773359962,0.7389572598424192,0.6611065536414182,0.8130154596146782 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weni/ZeroShot-3.4.22-Mistral-7b-DPO-1.0.0/results_2024-04-14T01-18-29.528053.json,Weni/ZeroShot-3.4.22-Mistral-7b-DPO-1.0.0,False,7248547840,0.894198481307534,0.7406309239618277,0.6566680057730363,0.8052786021889948 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weni/ZeroShot-Multilanguage-Zephyr-7B/results_2024-04-14T06-03-54.990451.json,Weni/ZeroShot-Multilanguage-Zephyr-7B,False,7268995072,0.8924749104951297,0.677400920042548,0.7136935903489231,0.8017592697770025 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weni/ZeroShot-Multilanguage-Zephyr-7B/results_2024-04-14T06-23-51.484845.json,Weni/ZeroShot-Multilanguage-Zephyr-7B,False,7268995072,0.8904227465531629,0.694821103041311,0.6194844910441241,0.7616189478881993 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weyaxi/Bagel-Hermes-34B-Slerp/results_2024-05-26T04-39-59.980925.json,Weyaxi/Bagel-Hermes-34B-Slerp,False,34388945920,0.9215534950233228,0.8223387188811104,0.837948717948718,0.7813796663968438 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weyaxi/Bagel-Hermes-34B-Slerp/results_2024-05-27T08-38-17.806381.json,Weyaxi/Bagel-Hermes-34B-Slerp,False,34388945920,0.9215534950233228,0.8223387188811104,0.837948717948718,0.7813796663968438 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weyaxi/Einstein-v7-Qwen2-7B/results_2024-07-16T19-16-29.654742.json,Weyaxi/Einstein-v7-Qwen2-7B,False,7615616512,0.9280457409761576,0.7719388373765231,0.7499744151659837,0.7412256806298099 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Weyaxi/Nous-Hermes-2-SUS-Chat-34B-Slerp/results_2024-06-16T06-32-30.997295.json,Weyaxi/Nous-Hermes-2-SUS-Chat-34B-Slerp,False,34388945920,0.9231939503620477,0.808342331307072,0.79766439567281,0.8489313135505292 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/WizardLMTeam/WizardLM-13B-V1.0/results_2024-05-26T18-03-45.219862.json,WizardLMTeam/WizardLM-13B-V1.0,False,13015874560,0.0,0.0,0.0,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/WizardLMTeam/WizardLM-70B-V1.0/results_2024-05-22T14-35-47.414856.json,WizardLMTeam/WizardLM-70B-V1.0,False,68976664576,0.9182742513146895,0.8237007597487587,0.7591403162055336,0.8563585077164766 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/X-D-Lab/XrayQwen/results_2024-09-22T03-01-01.507867.json,X-D-Lab/XrayQwen,False,9656935168,0.7951257593130199,0.6004986729914701,0.6090209456007647,0.458551298246319 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/xverse/XVERSE-13B/results_2024-02-13T10-06-05.666208.json,xverse/XVERSE-13B,False,13717652480,0.5320021613901151,0.11054715860040873,0.46760523254433656,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/xverse/XVERSE-13B/results_2024-04-02T16-25-23.716110.json,xverse/XVERSE-13B,False,13717652480,0.853651791779636,0.5612184700878184,0.495072463768116,0.7506150372392411 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/xverse/XVERSE-65B/results_2024-02-11T18-06-07.804136.json,xverse/XVERSE-65B,False,66408521728,0.7769915017384157,0.49474865786330113,0.4617741808753045,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/xverse/XVERSE-65B/results_2024-02-19T06-41-25.853474.json,xverse/XVERSE-65B,False,66408521728,0.8703715365817846,0.7327314288775599,0.4891375750750751,0.7725655170589092 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/xverse/XVERSE-65B-2/results_2024-02-12T02-54-57.964064.json,xverse/XVERSE-65B-2,False,66408521728,0.3421510912238783,0.09069952456959236,0.27583102825954914,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/xverse/XVERSE-65B-2/results_2024-03-04T21-29-41.764782.json,xverse/XVERSE-65B-2,False,66408521728,0.5908526162127826,0.74555509267116,0.2507362534948742,0.49961151567856854 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/xverse/XVERSE-7B/results_2024-02-13T08-04-57.866850.json,xverse/XVERSE-7B,False,7299846144,0.41898754643009994,0.11265884885309121,0.5536234888632623,0.0 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/xverse/XVERSE-7B/results_2024-04-02T15-44-20.542908.json,xverse/XVERSE-7B,False,7299846144,0.5490230059947407,0.32545061197760405,0.5673385012919897,0.8375068706778643 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Xwin-LM/Xwin-LM-13B-V0.1/results_2024-05-26T19-36-11.616364.json,Xwin-LM/Xwin-LM-13B-V0.1,False,13015864320,0.8542892080220075,0.6630481619048496,0.543859649122807,0.726020880366055 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Xwin-LM/Xwin-LM-13B-V0.2/results_2024-05-26T21-04-51.418260.json,Xwin-LM/Xwin-LM-13B-V0.2,False,13015874560,0.8192528206184226,0.6760624895433014,0.5138369483919222,0.7667810479361502 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Xwin-LM/Xwin-LM-7B-V0.1/results_2024-05-22T06-12-57.385680.json,Xwin-LM/Xwin-LM-7B-V0.1,False,6738415616,0.4608494235190403,0.41595452428802293,0.5889854112404209,0.7156285451893469 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/Xwin-LM/Xwin-LM-7B-V0.2/results_2024-05-22T05-22-42.120566.json,Xwin-LM/Xwin-LM-7B-V0.2,False,6738423808,0.5601802707468968,0.46987778700351723,0.5675144244615034,0.7615719118756412 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/yahma/llama-7b-hf/results_2024-05-29T00-35-46.426145.json,yahma/llama-7b-hf,False,6738415616,0.3864350929444941,0.15399203445886336,0.525274656970149,0.3807353552771385 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/YOYO-AI/Qwen2.5-32B-YOYO-reasoning/results_2025-07-08T01-32-55.142078.json,YOYO-AI/Qwen2.5-32B-YOYO-reasoning,False,32759790592,0.9460625544572304,0.7701536906039239,0.8293404606185044,0.9191946433505926 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/yunconglong/DARE_TIES_13B/results_2024-05-18T23-38-30.233980.json,yunconglong/DARE_TIES_13B,False,12879138816,0.9252440876057113,0.773934922061916,0.778378038405646,0.7573095601893222 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/yunconglong/MoE_13B_DPO/results_2024-07-04T01-29-22.264056.json,yunconglong/MoE_13B_DPO,False,12879138816,0.9256518082422204,0.7785639339764251,0.7736071481198676,0.754808735771493 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/yunconglong/Truthful_DPO_TomGrc_FusionNet_7Bx2_MoE_13B/results_2024-05-18T10-42-26.679204.json,yunconglong/Truthful_DPO_TomGrc_FusionNet_7Bx2_MoE_13B,False,12879138816,0.9256504186602871,0.7758187578008335,0.7697183994826833,0.7573095601893222 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/zetasepic/Qwen2.5-32B-Instruct-abliterated-v2/results_2025-07-10T10-08-12.952956.json,zetasepic/Qwen2.5-32B-Instruct-abliterated-v2,False,32763876352,0.9448359207492086,0.7458593183192128,0.8004230349344978,0.9285701165939783 | |
| /mnt/hgfs/D/dev/llm_pt_leaderboard_results/zhengr/MixTAO-7Bx2-MoE-v8.1/results_2024-05-17T19-51-43.191836.json,zhengr/MixTAO-7Bx2-MoE-v8.1,False,12879138816,0.9211590650705042,0.7751642830429548,0.7769607843137254,0.8200809063593439 | |